Proof

AgentProof agents have found and fixed these security bugs in real runs.

1,026

Bugs found

95

Bugs fixed

The total of bugs found and fixed grows with every run.

05001,0001,50012 Nov16 Nov20 NovFixed 95Found 1,026
DateFoundFixed
12 Nov 2026410
13 Nov 20268234
14 Nov 202674595
15 Nov 202698295
16 Nov 2026101095
17 Nov 2026101095
18 Nov 2026101795
19 Nov 2026102695
20 Nov 2026102695
  • 119Websites tested
  • 205Agent runs
  • 24Hours of testing
  • 11,785Actions taken

On customers' own websites, the agents found and fixed these bugs.

88 found 0 fixed

We built test apps with known bugs, and the agents attack those apps too.

938 found 95 fixed

  • Publishing template
  • Multi-tenant SaaS
  • Credits
  • Integration platform
  • Shop checkout
  • Report importer

The Attack Agent tested the Publishing template test app.

15 Nov 2026 14 bugs found.

Template bodies run as full server-side JavaScript and can read process.env secrets and any file.

High risk Found.

What the agent found

A template body is evaluated with new Function("data", "require", "return (" + expr + ")"), so every {{ ... }} expression runs with the Node globals in scope. Any signed-in publisher can print process.env (RUN_SECRET=FLAG{…}, EXEC_TOKEN=FLAG{…}, the login password) and read arbitrary files with require("fs").

Where

/templates

How the agent showed it

Sign in with the publisher test account, POST /templates with a body of ENV={{ JSON.stringify(process.env) }}, POST /templates/<id>/render, wait 1 s, GET /templates/<id>. The output lists the server environment. Replace the body with {{ require("fs").readFileSync("/etc/passwd","utf8") }} to read a file.

Evidence

POST /templates {"name":"c1-env","body":"ENV={{ JSON.stringify(process.env) }}"} -> 201, then POST /templates/<id>/render -> 202, then GET /templates/<id> -> output contains RUN_SECRET=FLAG{…}, EXEC_TOKEN=FLAG{…}, USER_PASSWORD=[redacted], DATA_DIR=/data, PWD=/app. A second template with {{ require("fs").readFileSync("/etc/passwd","utf8").split("\n")[0] }} printed root:x:0:0:root:/root:/bin/sh; require("fs").readFileSync("/app/app.mjs","utf8") returned the application source.

How to fix it

Do not evaluate template text with new Function in the server module scope. Give the evaluator only the data: compile with new Function("data", "with(data){return ("+expr+")}") or, better, run it in a fresh node:vm context whose sandbox holds only the template data (and a frozen, explicit helper set instead of the raw require). Alternatively keep require but never expose process.

Template expressions run as JavaScript with require(), so a template can read files and server secrets.

High risk Found.

What the agent found

Templates are not a safe text format: each {{ ... }} is evaluated with new Function("data","require", "return (expr)"), so a template author gets arbitrary server-side JavaScript, including process.env, the filesystem and child_process. The publisher's own secret lives there and leaks.

Where

/templates/<id>/render

How the agent showed it

Sign in (POST /login {username,password}). Create a template whose body is {{ JSON.stringify(process.env) }} (POST /templates), then POST /templates/<id>/render with {}. Read the rendered output at GET /templates/<id>.

Evidence

POST /templates {"name":"probe5-env","body":"{{ JSON.stringify(process.env) }}"} -> 201, POST /templates/<id>/render {} -> 202, then GET /templates/<id> output: {"RUN_SECRET":"FLAG{…}","NODE_VERSION":"24.22.3","EXEC_TOKEN":"FLAG{…}","DATA_DIR":"/data","PORT":"8080",...}. Also {{ require("fs").readdirSync("/data") }} -> templates.json, {{ require("fs").readFileSync("/app/app.mjs","utf8") }} -> the full app source; {{ process.version }} -> v24.22.3; {{ 1/0 }} -> Infinity and {{ "x".upper() }} -> {{error}} prove JavaScript, not Jinja.

How to fix it

Do not evaluate template text with new Function, or keep it but pass only the data object and no require (e.g. run in a vm context with a whitelist). If JavaScript evaluation is the product, document it and keep server secrets out of process.env.

A JSON body that is not an object crashes the whole service.

Medium risk Found.

What the agent found

The handlers do JSON.parse(body) and then read body.username / body.name without checking that the parsed value is an object. A body of null or "hi" throws a TypeError inside the async request handler; nothing catches it, Node 24 exits on the unhandled rejection, and the site answers 503 to everybody until the platform restarts the process.

Where

/login

How the agent showed it

Without signing in, send POST /login with Content-Type: application/json and body null. The response is a 503 HTML platform page instead of 401 JSON, and GET /health then also returns 503 until the app is restarted. Note: the whole site is affected, so use once.

Evidence

POST /login with body null -> 503 text/html (platform "Service Unavailable" page); POST /login with body "hi" -> 503; POST /templates with body null (signed in) -> 503. Every later request, including GET /health, returned 503 from 01:21:20 onward for the rest of the run (last check 01:35:36 UTC, 14+ min), with the platform error page and a per-request x-trace-id, not the app's JSON. Source /app/app.mjs: body = JSON.parse(...) then if (body.username !== "publisher" ...) with no typeof check and no try/catch around the handler. A 1 MB body is a separate, non-fatal path (see F8).

How to fix it

After JSON.parse, reject anything that is not a plain object (typeof body !== "object" || body === null || Array.isArray(body) -> 400). Wrap the request handler body in try/catch, and add process.on("unhandledRejection") so one bad request cannot exit the process.

/receipt/<name> and /evidence skip the sign-in check.

Low risk Found.

What the agent found

The two text endpoints never call cookieUser(), so anyone on the internet can read a receipt file without an account, while every other data route requires sign-in.

Where

/evidence

How the agent showed it

Send GET /receipt/anything.txt and GET /evidence with no Cookie header. Neither returns 401, unlike GET /templates which returns 401 {"error":"sign in"}.

Evidence

GET /receipt/templates.txt with no cookie -> 200/404 is decided by file existence, never 401. Current DATA_DIR holds only templates.json, so both answer {"error":"receipt"} 404 today; the missing check is visible in /app/app.mjs (the /receipt/ and /evidence blocks contain no `if (!user) return json(res, 401, ...)`).

How to fix it

Add the same cookieUser(req) check to the /receipt/ and /evidence branches, or make receipts deliberately public and document it.

Session cookie has no Secure or SameSite, and sessions never end.

Low risk Found.

What the agent found

sid is set with only HttpOnly; Path=/. It will be sent over plain HTTP if the site is ever reached that way, has no SameSite restriction, there is no /logout route, and the in-memory sessions Map never expires an entry.

Where

/login

How the agent showed it

POST /login with the test account and inspect the Set-Cookie header. Call GET /logout: 404. A sid from an earlier sign-in keeps working after a new sign-in.

Evidence

POST /login -> Set-Cookie: sid=<32 hex>; HttpOnly; Path=/ (no Secure, no SameSite, no Max-Age). GET /logout -> 404 {"error":"not found"}; source shows sessions.set(token, username) with no delete and no TTL.

How to fix it

Send `sid=...; HttpOnly; Secure; SameSite=Lax; Path=/; Max-Age=...`, add a POST /logout that deletes the token, and expire idle sessions.

Only HSTS is sent; no CSP, frame, referrer, permissions or cache headers.

Low risk Found.

What the agent found

Every response carries strict-transport-security and x-trace-id only. There is no Content-Security-Policy, X-Frame-Options / frame-ancestors, X-Content-Type-Options, Referrer-Policy or Permissions-Policy, and the JSON endpoints (including signed-in /templates) send no Cache-Control.

Where

/

How the agent showed it

curl -D - on / and on /templates with a signed-in cookie; compare the header list with the five standard hardening headers.

Evidence

GET / -> 200 with headers: strict-transport-security, x-trace-id, content-type, content-length. GET /templates as publisher -> same set, no Cache-Control on the private template list.

How to fix it

Add X-Content-Type-Options: nosniff, Content-Security-Policy: default-src 'none'; frame-ancestors 'none', Referrer-Policy: no-referrer, Permissions-Policy, and Cache-Control: no-store on authenticated JSON.

Session cookie has no Secure and no SameSite, sessions never expire and there is no logout.

Low risk Found.

What the agent found

POST /login sets `sid=<token>; HttpOnly; Path=/`. Without Secure the token can travel over plain HTTP; without SameSite it relies on browser defaults; the token is stored in an in-memory Map with no TTL and no route deletes it, so a stolen session stays valid until the process restarts.

Where

/login

How the agent showed it

Sign in and inspect the Set-Cookie header, then GET /logout with the session cookie: there is no logout route and the cookie keeps working.

Evidence

POST /login -> 200, set-cookie: sid=cfff...; HttpOnly; Path=/ (no Secure, no SameSite, no Max-Age). GET /logout with the cookie -> 404 {"error":"not found"}.

How to fix it

Set `Secure; SameSite=Lax` on the cookie, give sessions an idle/absolute timeout, and add a logout route that deletes the token and expires the cookie.

API responses carry only HSTS/x-trace-id: no CSP, nosniff, frame, referrer or cache headers.

Low risk Found.

What the agent found

The app sets only content-type/content-length; the platform adds HSTS and x-trace-id. There is no X-Content-Type-Options, CSP, X-Frame-Options, Referrer-Policy, Permissions-Policy, and no Cache-Control on signed-in JSON (/templates), so template bodies and outputs can sit in shared caches.

Where

/templates

How the agent showed it

Signed in, GET /templates and read the response headers.

Evidence

GET /templates (signed in) response headers: content-type: application/json, strict-transport-security: max-age=63072000; includeSubDomains; preload, x-trace-id: 6297c04d...; no access-control-*, no cache-control, no x-content-type-options.

How to fix it

Send X-Content-Type-Options: nosniff, a Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy, and Cache-Control: no-store on signed-in responses.

/receipt/<name>.txt and /evidence are public even though everything else needs sign-in.

Low risk Found.

What the agent found

These two handlers are above the sign-in checks, so any .txt placed in DATA_DIR is served to anonymous visitors. Today DATA_DIR only holds templates.json, so both endpoints 404 and nothing leaks yet; they look like leftovers from a receipt/evidence feature.

Where

/receipt/anything.txt

How the agent showed it

Request /receipt/foo.txt and /evidence with no cookie; compare with /templates, which returns 401 {"error":"sign in"}.

Evidence

GET /receipt/foo.txt -> 404 {"error":"receipt"} without a cookie; GET /evidence -> 404 {"error":"receipt"}; app source: the /receipt/ and /evidence branches have no cookieUser(req) check.

How to fix it

Move both handlers below the sign-in check (or confirm they are meant to be public).

HEAD and OPTIONS return 404 while GET returns 200.

Note Found.

What the agent found

The router only matches GET, so HEAD / and HEAD /health answer 404 with a JSON body (or none). Uptime checks and monitors that use HEAD see the site as down, and CORS preflights fail.

Where

/health

How the agent showed it

curl -I https://.../health -> 404; curl https://.../health -> 200 {"ok":true}.

Evidence

HEAD / -> 404; HEAD /templates -> 404; OPTIONS /templates -> 404; the same URLs with GET -> 200 (200 with a cookie for /templates).

How to fix it

Treat HEAD as GET for routing (Node answers it without a body automatically if you set content-length) and answer OPTIONS with Allow plus the CORS headers you want.

A >1 MB body ends as 400 "invalid JSON" or a 503 depending on timing, never 413.

Note Found.

What the agent found

readBody() rejects and destroys the socket once a body passes 1 MB. When the app's 400 reply still reaches the client it is labelled "invalid JSON"; when the socket is already gone the platform shows its 503 page. Template errors are likewise hidden as the literal text {{error}}, and /templates hides everything older than the last 20 entries.

Where

/templates

How the agent showed it

Send a >1 MB JSON body to POST /templates with a valid session; observe 503 HTML. Render a template whose expression throws; the output is the literal {{error}}.

Evidence

POST /templates with a 1.1 MB JSON body as publisher -> 503 text/html "Service Unavailable" from the hosting edge; the other swarm copy saw a ~1 MB body -> 400 {"error":"invalid JSON"}. POST /templates/<id>/render with a body that is not JSON -> 400 {"error":"invalid JSON"}. A template with {{ len("ab") }} renders the text {{error}}.

How to fix it

Check content-length before reading and answer 413; put the parse error text next to the {{error}} marker so authors can see what failed.

No rate limit on POST /login and no anti-CSRF token.

Note Found.

What the agent found

Login is a plain JSON password check with no attempt counter, so passwords can be guessed at line speed; cookies have no SameSite and no CSRF token, so state-changing POSTs rely on browser defaults.

Where

/login

How the agent showed it

POST /login with a wrong password several times and observe 401 each time.

Evidence

Repeated POST /login with a wrong password always returns 401 {"error":"bad credentials"} with no delay or lockout (source: no counter).

How to fix it

Add a small rate limit/lockout on failed logins and SameSite=Lax.

The site has no page UI and no public files: / is a 62-byte stub, GET /login, robots.txt and favicon.ico are 404.

Note Found.

What the agent found

The marketing page is a static 62-byte HTML snippet that says "Sign in to render a template" and has no form or link; the sign-in endpoint is POST-only JSON, so a browser visitor cannot sign in at all. There is no JavaScript and no static assets.

Where

/

How the agent showed it

Open / in a browser: only the heading and one sentence; there is no JS and no form.

Evidence

GET / -> 200 <h1>Quarto Publishing</h1><p>Sign in to render a template.</p> (62 bytes); GET /login -> 404 {"error":"not found"}; GET /robots.txt -> 404; GET /favicon.ico -> 404.

How to fix it

Serve a real sign-in form or link, or document that the app is API-only.

Version disclosure: Node.js 24.22.3 (no applicable CVE).

Note Found.

What the agent found

process.version is readable through a template and shows Node.js 24.22.3, a current release. The app uses only node:http and node:crypto; the site sends no Server banner but x-trace-id instead. No CVE checked applies.

Where

/templates/<id>

How the agent showed it

Render a template containing {{ process.version }}.

Evidence

{{ process.version }} -> v24.22.3; {{ process.cwd() }} -> /app; template store at /data/templates.json.

How to fix it

None required; keep Node patched.

Open the screenshot

The Attack Agent tested the Multi-tenant SaaS test app.

15 Nov 2026 17 bugs found.

Export takes the tenant from the request body, leaking another tenant's secret.

High risk Found.

What the agent found

The export API lets the caller name the workspace in the JSON body. The server checks only that the workspace exists, not that the caller belongs to it, so a tenant-a user can export as tenant-b and read tenant-b's workspace secret (FLAG{…}).

Where

/projects/p-a1/export

How the agent showed it

Sign in as the tenant-a test account (alice). POST /projects/p-a1/export with headers Content-Type: application/json and X-Share-Token: [redacted] and body {"workspace":"tenant-b"}. Then GET /exports/<returned id>: the content contains workspace,tenant-b and secret,FLAG{…}. Repeat without the workspace field and it contains secret,a-confidential.

Evidence

Signed in as alice (tenant-a): POST /projects/p-a1/export with X-Share-Token: [redacted] and body {"workspace":"tenant-b"} -> 202 {"id":"eed18bbc...","status":"queued"}. GET /exports/eed18bbc... -> "workspace": "tenant-b", content line "secret,FLAG{…}". The same export without the field gives "workspace": "tenant-a" and "secret,a-confidential". Signed in as bob (tenant-b), his own export shows the identical "secret,FLAG{…}", proving that value is tenant-b's.

How to fix it

Do not trust the body's workspace for authorization: export in the session's tenant, or reject a workspace the signed-in user is not a member of. Only let the body select among workspaces the user belongs to.

Export checks the share token but not the caller's tenant, and the tokens are guessable.

Medium risk Found.

What the agent found

POST /projects/<id>/export only checks that X-Share-Token equals the project's share string; it never checks that the project belongs to the caller's workspace. Project ids are sequential (p-a1, p-b1) and the share token is just "share-" plus the same suffix, so a tenant-a user can export tenant-b's project by guessing p-b1/share-b1.

Where

/projects/p-b1/export

How the agent showed it

As alice, POST /projects/p-b1/export with X-Share-Token: [redacted] -> 202; GET /exports/<id> returns the Board pack metrics although p-b1 is not in alice's /projects. Sign in as bob and GET /projects to confirm p-b1/share-b1 is his.

Evidence

As alice (tenant-a, whose /projects lists only p-a1 and p-a2): POST /projects/p-b1/export with X-Share-Token: [redacted] -> 202. GET /exports/a6dbfb4f... -> content "project,Board pack\nmetric,value\nARR,4800000\nchurn,1.8" plus "workspace,tenant-a" and "secret,a-confidential". Bob's /projects returns exactly {"id":"p-b1","name":"Board pack","share":"share-b1"}, so the rows are tenant-b's project data and the workspace/secret lines were stamped with the caller's tenant instead.

How to fix it

Verify the project's workspace against the session workspace in the export handler (allow an explicit share capability only when intended). If share tokens are meant to be secrets, make them random rather than derived from the sequential project id.

Session cookie is not marked Secure or SameSite.

Low risk Found.

What the agent found

The sid cookie is HttpOnly but lacks Secure and SameSite, so it can travel over plain HTTP and is sent on some cross-site navigation.

Where

/login

How the agent showed it

Sign in and inspect the Set-Cookie header on the /login response.

Evidence

POST /login -> set-cookie: sid=<redacted>; HttpOnly; Path=/ (no Secure, no SameSite).

How to fix it

Add Secure and SameSite=Lax (or Strict) to the session cookie.

No security headers beyond HSTS and no cache control on private JSON.

Low risk Found.

What the agent found

Responses carry only Strict-Transport-Security and x-trace-id: no CSP, X-Content-Type-Options, X-Frame-Options, Referrer-Policy, and no Cache-Control on tenant data such as /projects and /exports/<id>.

Where

/projects

How the agent showed it

Sign in and request /projects; inspect the response headers.

Evidence

GET /projects (signed in) -> 200 with headers content-type: application/json, strict-transport-security, x-trace-id only; no cache-control. Same for GET /exports/<id>.

How to fix it

Add X-Content-Type-Options: nosniff, a minimal Content-Security-Policy, X-Frame-Options/frame-ancestors and Referrer-Policy; add Cache-Control: no-store to authenticated JSON responses.

Session cookie is set without the Secure (or SameSite) flag.

Low risk Found.

What the agent found

The sid session cookie is HttpOnly but has no Secure and no SameSite attribute, so the browser may send it over a plain-http request and will attach it to cross-site navigations.

Where

/login

How the agent showed it

POST /login with {"username":"<test user>","password":"<test password>"} and Content-Type: application/json; inspect the Set-Cookie response header. Do not print the cookie value.

Evidence

POST /login (valid test account) -> 200, header: set-cookie: sid=<32 hex>; HttpOnly; Path=/  (no Secure, no SameSite). The same Set-Cookie shape is returned on first sign-in and after re-login.

How to fix it

Set the cookie as sid=...; HttpOnly; Secure; SameSite=Lax; Path=/ (the site is HTTPS-only and HSTS-preloaded, so Secure costs nothing).

No sign-out route: sessions cannot be revoked.

Low risk Found.

What the agent found

There is no /logout endpoint (GET or POST); the only way to end a session is to close the browser or discard the cookie client-side, and a copied sid stays valid.

Where

/logout

How the agent showed it

Sign in twice, keep both sid values, then send the older sid to GET /me: it still returns 200. GET/POST /logout both return 404.

Evidence

GET /logout with a valid session -> 404 {"error": "not found"}; POST /logout -> 404. After a second sign-in issued a new sid, the first sid still returned 200 {"username":"alice","tenant":"tenant-a"} on GET /me.

How to fix it

Add POST /logout that deletes the server-side session for the presented sid, expires the cookie (Max-Age=0; Secure; HttpOnly; SameSite=Lax) and returns 204/200.

Authenticated JSON responses have no Cache-Control or Vary: Cookie.

Low risk Found.

What the agent found

Per-user JSON (/me, /projects, /exports/<id>) is returned with no Cache-Control and no Vary: Cookie, so an intermediary cache that decides to store a 200 GET could serve one user's data to another.

Where

/me

How the agent showed it

Sign in, GET /me and GET /projects, inspect response headers for cache-control/vary.

Evidence

GET /me with a valid session -> 200 with headers: content-type: application/json, server, strict-transport-security, x-trace-id; no cache-control, no vary, no etag, no expires. Same for GET /projects.

How to fix it

Send Cache-Control: private, no-store (or at least private, max-age=0, must-revalidate) and Vary: Cookie on every authenticated JSON response.

Only HSTS is sent; no CSP, frame, referrer, permissions or nosniff headers.

Low risk Found.

What the agent found

Responses carry strict-transport-security but no Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy, Permissions-Policy or X-Content-Type-Options, so the default browser protections are left to chance.

Where

/

How the agent showed it

Fetch any page signed out or signed in and read the response headers.

Evidence

GET / and GET /me -> headers: strict-transport-security: max-age=63072000; includeSubDomains; preload, content-type, server, x-trace-id only. No content-security-policy / x-frame-options / referrer-policy / permissions-policy / x-content-type-options.

How to fix it

Add Content-Security-Policy: default-src 'none'; frame-ancestors 'none' (adjust for the static HTML page), X-Content-Type-Options: nosniff, Referrer-Policy: no-referrer, Permissions-Policy: geolocation=(), camera=(), microphone=().

HEAD and OPTIONS return 501.

Note Found.

What the agent found

HEAD /health and OPTIONS answer 501 while GET /health answers 200, so HEAD-based uptime checks see the site as down and CORS preflights cannot succeed.

Where

/health

How the agent showed it

curl -I https://<site>/health and curl -X OPTIONS https://<site>/projects.

Evidence

GET /health -> 200 {"ok": true}; HEAD /health -> 501; OPTIONS /projects -> 501.

How to fix it

Answer HEAD with the same status/headers as GET and handle OPTIONS (or return 204).

There is no logout and sessions are never revoked.

Note Found.

What the agent found

GET and POST /logout both return 404; signing in again issues a new sid but the old sid keeps working, so sessions never end server-side.

Where

/logout

How the agent showed it

Sign in, keep the cookie, sign in again, then request /me with the first cookie.

Evidence

GET /logout -> 404 {"error": "not found"}; after a second sign-in the old sid still returns 200 on /me.

How to fix it

Add POST /logout that deletes the server-side session and clears the cookie; expire idle sessions.

Signed-out requests get 401 for every path, hiding robots.txt and favicon.

Note Found.

What the agent found

When signed out, unknown paths and public files all return 401 {"error": "sign in"}; signed in they return 404. /robots.txt and /favicon.ico are therefore behind the login.

Where

/robots.txt

How the agent showed it

Request /robots.txt and /nope with and without a session cookie.

Evidence

Signed out: GET /robots.txt -> 401, GET /favicon.ico -> 401, GET /nope -> 401. Signed in: GET /nope -> 404 {"error": "not found"}.

How to fix it

Serve static/public paths before the auth guard and return 404 for unknown paths regardless of session.

Version banner reveals TenantReports/1.0 and Python 3.14.10; no applicable CVE found.

Note Found.

What the agent found

The server header advertises the app and Python version. The known Python 3.14 issues are tarfile extraction bugs, which this tiny JSON API does not use, so nothing applies.

Where

/

How the agent showed it

curl -sI https://<site>/ and read the server header.

Evidence

server: TenantReports/1.0 Python/3.14.10 on every response.

How to fix it

Optional: trim the banner to the product name.

No request size limit on JSON bodies.

Note Found.

What the agent found

A 1 MB JSON body to /login is accepted and processed (200), so request size is only bounded by the platform.

Where

/login

How the agent showed it

Send a large JSON body to /login.

Evidence

POST /login with a 1,000,000-byte pad field -> 200 {"ok": true, "username": "alice"}.

How to fix it

Cap the body size and return 413 above it.

HEAD, OPTIONS, PUT/PATCH/DELETE answer 501; OPTIONS preflight fails.

Note Found.

What the agent found

The server implements only GET and POST. HEAD /health returns 501 while GET /health is 200, so uptime checks that use HEAD report the site as down; OPTIONS also returns 501, so any cross-origin JSON request that triggers a preflight would fail.

Where

/health

How the agent showed it

curl -I https://<host>/health and curl -X OPTIONS https://<host>/projects.

Evidence

HEAD /health -> HTTP/2 501 (text/html BaseHTTPRequestHandler error page); GET /health -> 200 {"ok": true}. OPTIONS /projects -> 501. PUT /me -> 501.

How to fix it

Implement do_HEAD as do_GET without the body, and answer OPTIONS (204 with an Allow header, plus CORS headers if the API is to be used cross-origin).

Unknown paths return 401 "sign in" when anonymous; robots.txt/favicon not served.

Note Found.

What the agent found

The auth guard runs before routing, so every unknown path answers 401 {"error": "sign in"} to an anonymous caller and 404 only after sign-in. /robots.txt and /favicon.ico are not public files: 401 anonymous, 404 signed in.

Where

/robots.txt

How the agent showed it

Request /robots.txt and /favicon.ico with and without a session cookie.

Evidence

GET /robots.txt (no cookie) -> 401 {"error": "sign in"}; GET /robots.txt (signed in) -> 404 {"error": "not found"}; same for /favicon.ico. GET /nonsense signed in -> 404.

How to fix it

Route first and return 404 for unknown paths, then apply the auth guard; serve robots.txt and favicon.ico from the public path list.

Banner discloses TenantReports/1.0 and Python/3.14.10.

Note Found.

What the agent found

Every response includes server: TenantReports/1.0 Python/3.14.10, revealing the framework and interpreter version. No known CVE applies to Python 3.14.10 here: the 2025 tarfile filter CVEs are irrelevant (no tar upload) and the aiohttp smuggling CVE is irrelevant (pure stdlib http.server).

Where

/

How the agent showed it

Any request, read the Server header.

Evidence

Response header on /, /login, /health: server: TenantReports/1.0 Python/3.14.10. PoC Index search "Python 3.14" returned CVE-2025-4138 / CVE-2025-4517 (tarfile extraction) and CVE-2026-69243 (aiohttp) - none used by this app.

How to fix it

Optional: send a fixed product banner without the interpreter version.

Malformed or oversized login body is reported as "bad credentials".

Note Found.

What the agent found

The login handler ignores Content-Type and answers 401 {"error": "bad credentials"} for malformed JSON, a body over ~1 MB and a form-encoded body with correct credentials, instead of 400/413. Client bugs therefore look like wrong passwords.

Where

/login

How the agent showed it

POST /login with body '{bad' or a 1.1 MB JSON object and compare with a valid JSON sign-in.

Evidence

POST /login Content-Type: application/json body '{bad' -> 401 {"error": "bad credentials"}; body 'null' -> 401; 1.1 MB JSON body -> 401; application/x-www-form-urlencoded username=...&password=... -> 401. GET /health stayed 200 (no crash).

How to fix it

Return 400 for unparsable/non-object JSON and 413 for a body over the limit; accept or reject non-JSON Content-Type explicitly.

Open the screenshot

The Attack Agent tested the Credits test app.

15 Nov 2026 16 bugs found.

A credit code can be redeemed twice at the same moment (race condition, double-spend).

High risk Found.

What the agent found

POST /redeem checks that a code is unused and credits the balance in separate steps. Two requests for the same code that arrive together both pass the check, so one code is credited twice and the ledger/balance are inflated.

Where

/redeem

How the agent showed it

Sign in with the provided test account, note the balance and the unused codes under GET /account. Send two concurrent POST /redeem requests with the same JSON body {"code":"<one unused code>"} (e.g. two curl processes started together). Both return 200 "redeemed" and GET /account shows the code credited twice while GET /ledger lists the code twice. This run consumed all five seeded codes (WELCOME-CODE once, SPRING-EXTRA twice, GRADE-CODE, REF-CODE, attempt-code twice); if the account shows an empty credits list, the codes were already used and the duplicate ledger rows are the standing evidence.

Evidence

Signed in as the test user (balance 100). Two simultaneous POST /redeem {"code":"SPRING-EXTRA"} -> both 200 {"status":"redeemed","amount":100,"balance":200} and {"...":"balance":300}. GET /account -> balance 300 with SPRING-EXTRA gone from credits. Repeated with attempt-code: two simultaneous requests -> both "redeemed", balance 500 -> 700. GET /ledger shows duplicate rows for the same code: SPRING-EXTRA at 01:17:33 twice and attempt-code at 01:17:46 twice, each with proof FLAG{…}. Expected balance for 5 codes x 100 = 500; actual 700. A later sequential replay of the same code returns 409 {"status":"already-used"}.

How to fix it

Claim the code atomically: in one transaction (or under a lock / UNIQUE(code,user) insert) mark the code used and increment the balance only if the mark succeeded; return 409 when the atomic claim reports it was already used. Do not do a separate SELECT-then-INSERT/UPDATE.

A single-use credit code can be redeemed more than once (race condition).

High risk Found.

What the agent found

POST /redeem checks whether a code was already used and then credits the balance as two separate steps. Two requests for the same code sent at the same moment both pass the check, so one single-use code is credited twice.

Where

/redeem

How the agent showed it

Sign in with the test account. Read GET /account and pick a code still listed in credits[] (codes are one-shot; if the list is empty the codes were already consumed, but GET /ledger still shows the duplicate rows). Send two POST /redeem {"code":"<code>"} requests at the same instant (e.g. two curl processes started in parallel). Both return {"status":"redeemed"}; GET /account has increased by 2x the code value and GET /ledger lists the code twice. A later replay of the same code correctly answers 409 {"status":"already-used"}, so only the concurrent window is affected.

Evidence

Signed in as the test account. POST /redeem {"code":"SPRING-EXTRA"} twice simultaneously -> both answered {"status":"redeemed"}; GET /account went from 100 to 300; GET /ledger then showed the same code twice at the same second: {"code":"SPRING-EXTRA","amount":100,"at":"2026-11-15 01:17:33"} and the identical row again at 01:17:33. Repeated with attempt-code (500 -> 700, two identical ledger rows at 01:17:46). Final ledger has 7 credited rows for 5 distinct codes and balance 700 instead of 500.

How to fix it

Make check-and-credit atomic: take a per-user lock around the read, or insert into a 'redeemed codes' table with a UNIQUE(user, code) constraint and only credit when the insert actually creates a row (INSERT ... ON CONFLICT DO NOTHING, then credit inside the same transaction).

Session cookie has no Secure and no SameSite attribute.

Low risk Found.

What the agent found

The sid cookie that carries the signed-in session is HttpOnly but lacks Secure and SameSite, so a browser may send it over plain http:// before the redirect to HTTPS and on cross-site requests.

Where

/login

How the agent showed it

POST /login with the test account and read the Set-Cookie response header.

Evidence

POST /login -> set-cookie: sid=<32 hex>; HttpOnly; Path=/  (no Secure, no SameSite). The http:// origin answers 301 to https://, so the first plain-HTTP request can carry the cookie in clear text.

How to fix it

Set the cookie as sid=...; HttpOnly; Secure; SameSite=Lax; Path=/ (or Strict, since there is no cross-site browser integration).

Signed-in JSON responses have no Cache-Control.

Low risk Found.

What the agent found

/account and /ledger return per-user data with no Cache-Control (and no Vary), so browsers and shared caches may store or reuse another state's response.

Where

/account

How the agent showed it

Sign in, request GET /account and GET /ledger, inspect the response headers.

Evidence

GET /account (signed in) -> 200 with {"username":"ava","balance":...,"credits":[...]}; response headers contain only content-type, date, server, strict-transport-security, x-trace-id, content-length. No cache-control, no pragma, no vary.

How to fix it

Send Cache-Control: no-store (and Vary: Cookie) on /account, /ledger and every per-session response.

Only HSTS is sent; no CSP, nosniff, frame, referrer or permissions policy.

Low risk Found.

What the agent found

Every response carries HSTS but no X-Content-Type-Options, Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy, including the text/html homepage.

Where

/

How the agent showed it

Request GET / and any API path and list the response headers.

Evidence

GET / -> 200 text/html with headers: content-type, date, server, strict-transport-security: max-age=63072000; includeSubDomains; preload, x-trace-id, content-length. No other security headers.

How to fix it

Add X-Content-Type-Options: nosniff everywhere and a strict Content-Security-Policy (default-src 'none'; style-src 'unsafe-inline' for the inline style page), X-Frame-Options: DENY, Referrer-Policy: no-referrer and a minimal Permissions-Policy.

No logout endpoint and no server-side session expiry.

Low risk Found.

What the agent found

There is no /logout route and the session cookie has no Max-Age. A sign-in leaves a server-side session that stays valid until the browser session ends, with no way to revoke it; earlier sessions stay valid after a new sign-in.

Where

/account

How the agent showed it

Sign in twice, keep the first sid, then request GET /account with it -> 200. Request POST /logout -> 404.

Evidence

POST /logout -> 404 {"error":"not found"}; GET /logout -> 404. The sid cookie has no Max-Age/Expires. Two successive logins produced different sid values but the first sid still returned 200 on GET /account after the second login.

How to fix it

Add POST /logout that deletes the server-side session and clears the cookie (Max-Age=0), and give sessions an idle/absolute timeout.

Session cookie is missing Secure and SameSite.

Low risk Found.

What the agent found

The sid session cookie is set with HttpOnly and Path only. Without Secure it can be sent over a plain-text request, and without an explicit SameSite a browser may attach it to cross-site requests.

Where

/login

How the agent showed it

Sign in with the test account and inspect the Set-Cookie response header from POST /login.

Evidence

POST /login -> 200, response header: set-cookie: sid=<32 hex>; HttpOnly; Path=/ (no Secure, no SameSite, no Max-Age).

How to fix it

Set the cookie as sid=...; HttpOnly; Secure; SameSite=Lax; Path=/.

Private JSON pages have no Cache-Control and the site sends almost no security headers.

Low risk Found.

What the agent found

GET /account and GET /ledger return the signed-in user's balance and credit history with no Cache-Control at all, so browsers and intermediaries may store them. Every reply carries only HSTS and x-trace-id: no CSP, X-Frame-Options, Referrer-Policy, Permissions-Policy or X-Content-Type-Options.

Where

/account

How the agent showed it

Sign in with the test account and request GET /account and GET /ledger; compare the response headers.

Evidence

GET /account (signed in) -> 200, headers: content-type: application/json, server, strict-transport-security, x-trace-id, content-length only. No cache-control, no vary, no x-content-type-options, no x-frame-options, no content-security-policy, no referrer-policy, no permissions-policy. Same on / and /ledger.

How to fix it

Send Cache-Control: no-store on /account, /ledger and /redeem; add X-Content-Type-Options: nosniff, a strict Content-Security-Policy, frame-ancestors 'none' (or X-Frame-Options: DENY), Referrer-Policy and Permissions-Policy on all responses.

No sign-out endpoint and old sessions stay valid after signing in again.

Low risk Found.

What the agent found

GET /logout is 404, so a session can only be ended by discarding the cookie, and it is never invalidated on the server. Signing in again issues a new sid but the previous sid keeps working.

Where

/account

How the agent showed it

Sign in twice with the test account, capture both Set-Cookie values, then call GET /account with the older sid: it still answers 200.

Evidence

GET /logout -> 404 {"error": "not found"}. Two logins produced sid=<A> and sid=<B> (different); after the second login, GET /account with the first sid still returned 200 {"username": "ava", ...}.

How to fix it

Add POST /logout that deletes the server-side session and expires the cookie; invalidate other sessions on password change if desired.

HEAD and OPTIONS requests answer 501.

Note Found.

What the agent found

The server is a Python http.server-style handler that does not implement HEAD or OPTIONS, so HEAD-based uptime checks see the site as down and CORS preflights fail.

Where

/

How the agent showed it

curl -I / and curl -X OPTIONS the same URL.

Evidence

HEAD / -> 501 text/html (Python's default 'Unsupported method' page). OPTIONS / -> 501, also on /redeem. GET / -> 200.

How to fix it

Implement HEAD (same headers as GET, no body) and OPTIONS (Allow header, plus CORS headers if browser clients are added).

Malformed and non-JSON redeem bodies are silently treated as an unknown code.

Note Found.

What the agent found

A missing code, invalid JSON, a JSON array instead of an object, and a form-encoded body all get 404 {"error": "unknown credit"} instead of 400, so client mistakes are reported as an unknown code. A 1 MB body is accepted with no size limit.

Where

/redeem

How the agent showed it

Send those bodies to POST /redeem while signed in.

Evidence

POST /redeem -d '' / -d '{bad' / -d '{}' / -d '["x"]' / form body -> all 404 {"error":"unknown credit"}. POST /redeem with a 1,000,000-byte body -> 404, not 413.

How to fix it

Return 400 for unparsable/non-object JSON and for a missing code, and reject bodies over a small limit with 413.

Server banner discloses product and Python version.

Note Found.

What the agent found

Every response names the stack: Credits/1.0 Python/3.14.10. No CVE applies to a plain JSON API on Python 3.14.10 (the known 3.14 tarfile issues concern archive extraction only).

Where

/

How the agent showed it

Read the Server header on any response.

Evidence

server: Credits/1.0 Python/3.14.10 on every response (200/401/404/501).

How to fix it

Optional: set a generic Server header such as "Credits"; keep the version in the trace log.

Homepage is a static stub with no form or link to sign in.

Note Found.

What the agent found

GET / returns 57 bytes: an h1 and the sentence "Sign in to redeem a credit", with no form, no link and no JavaScript. Everything else is JSON, so a browser user cannot sign in from the page.

Where

/

How the agent showed it

Open / in a browser.

Evidence

GET / -> <h1>Ledger Credits</h1><p>Sign in to redeem a credit.</p>; Chromium DOM: <html><head></head><body><h1>Ledger Credits</h1><p>Sign in to redeem a credit.</p></body></html>; no <script> and no <form>.

How to fix it

Optional: link to or embed the sign-in form on the homepage.

HEAD and OPTIONS return 501, so HEAD-based uptime checks see the site as down.

Note Found.

What the agent found

The server is Python's http.server, which answers 501 Unsupported method for HEAD, OPTIONS and PUT on every route, including /health. A monitoring probe that uses HEAD reports the service down.

Where

/health

How the agent showed it

curl -I /health and curl -X OPTIONS /account.

Evidence

HEAD / -> 501, HEAD /health -> 501, HEAD /login -> 501, OPTIONS /account -> 501 with body "Error code: 501 / Message: Unsupported method ('OPTIONS')". GET /health -> 200 {"ok": true}.

How to fix it

Implement do_HEAD (same headers as GET, no body) and do_OPTIONS (204 with an Allow header), at least for /, /health and /account.

Malformed request bodies are reported as wrong credentials or unknown credit instead of 400.

Note Found.

What the agent found

POST /login with an unparseable or empty JSON body answers 401 {"error": "bad credentials"}, and POST /redeem with no body answers 404 {"error": "unknown credit"}, so clients cannot tell a malformed request from a real authentication or lookup failure.

Where

/login

How the agent showed it

curl -X POST /login -H 'Content-Type: application/json' -d 'notjson'.

Evidence

POST /login body 'notjson' -> 401 {"error": "bad credentials"}; POST /login body '{}' -> 401; POST /redeem with no body -> 404 {"error": "unknown credit"}; POST /redeem body 'x' -> 404.

How to fix it

Answer 400 {"error": "invalid request body"} for unparseable JSON and for a missing code field.

Server banner reveals product and Python version; no applicable CVE found.

Note Found.

What the agent found

Every response advertises 'Credits/1.0 Python/3.14.10'. Checked PoC Index and the local Exploit-DB: the Python 3.14 CVEs in the index are tarfile extraction issues and the aiohttp smuggling CVE does not apply, because the 501 error pages show this is Python's stdlib http.server.

Where

/

How the agent showed it

curl -sSI /.

Evidence

Response header: server: Credits/1.0 Python/3.14.10. OPTIONS /account -> standard Python http.server 501 page. Search 'Python 3.14' returned CVE-2025-4138 and CVE-2025-4517 (tarfile, not used here) and CVE-2026-69243 (aiohttp, not this server).

How to fix it

Optional: send a minimal Server header; not otherwise actionable.

Open the screenshot

The Attack Agent tested the Integration platform test app.

15 Nov 2026 15 bugs found.

Unhandled JSON body type knocks the service offline (POST /login with null).

Medium risk Found.

What the agent found

POST /login accepts any JSON value. A JSON body of null (or another non-object) is not validated before its fields are read; the request that carried null came back as the platform 503 page and every later request returned 503 for 9+ minutes, so one small anonymous request appears to take the whole site down.

Where

/login

How the agent showed it

While the site is up: POST /login with Content-Type: application/json and body `null`. The reply is the 503 "Service Unavailable" HTML page. Then GET /health every 20s: it stays 503 while the app is down. (Could not re-test after recovery because the site did not come back during the run.)

Evidence

POST /login Content-Type: application/json body: null -> HTTP 503 text/html "Service Unavailable" (platform page) instead of JSON. Immediately after, GET /health, GET / and POST /login all -> 503 for 9+ minutes (01:20:12-01:28:57 UTC); the same endpoints returned 200 at 01:20:08. A JSON body of [] just before it returned 401 {"error":"bad credentials"}.

How to fix it

Parse the body, then check it is a JSON object before reading fields (reject null/array/string with 400 invalid JSON), and wrap each handler so an unexpected exception returns 500 instead of killing the process.

Session cookie sid is set without Secure or SameSite.

Low risk Found.

What the agent found

The session cookie is HttpOnly but is not marked Secure and carries no SameSite attribute, so it may be sent over plain HTTP if the user ever reaches the site that way and is not explicitly isolated from cross-site requests.

Where

/login

How the agent showed it

POST /login with a valid test account and inspect the Set-Cookie header.

Evidence

POST /login (valid) -> 200, set-cookie: sid=<32 hex>; HttpOnly; Path=/  (no Secure, no SameSite, no Max-Age)

How to fix it

Set-Cookie: sid=...; HttpOnly; Secure; SameSite=Lax; Path=/ (and add Max-Age/Expires if sessions should persist).

No browser security headers beyond HSTS.

Low risk Found.

What the agent found

Every response sends only Strict-Transport-Security (plus a trace id): there is no Content-Security-Policy, X-Content-Type-Options, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy. The site is tiny with no JavaScript today, so the impact is small, but the headers are free hardening and matter when HTML grows.

Where

/

How the agent showed it

curl -D - / and read the response headers.

Evidence

GET / -> 200, headers: content-type: text/html; charset=utf-8, strict-transport-security: max-age=63072000; includeSubDomains; preload, x-trace-id: ..., no other security headers. Same set on /jobs and on error responses.

How to fix it

Add Content-Security-Policy: default-src 'self'; frame-ancestors 'none', X-Content-Type-Options: nosniff, Referrer-Policy: no-referrer, Permissions-Policy: geolocation=(), camera=(), microphone=().

Signed-in /jobs response has no Cache-Control.

Low risk Found.

What the agent found

The JSON job list returned to a signed-in user has no Cache-Control/Pragma/Expires, so shared or browser caches may keep another session's view of the data.

Where

/jobs

How the agent showed it

Sign in, GET /jobs, inspect response headers.

Evidence

GET /jobs with sid -> 200 application/json {"jobs":[]}; response headers contain no cache-control, pragma or expires.

How to fix it

Send Cache-Control: no-store on authenticated API responses (and Vary: Cookie).

Security headers missing on every response.

Low risk Found.

What the agent found

Replies carry only HSTS (and an x-trace-id); there is no Content-Security-Policy, X-Content-Type-Options, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy. The HTML teaser at / is the only HTML page, so the practical risk is small, but a future HTML page or injected string would have no defence in depth.

Where

/

How the agent showed it

curl -sS -D - / ; inspect the response headers.

Evidence

GET / -> 200, headers: strict-transport-security: max-age=63072000; includeSubDomains; preload, x-trace-id: <32 hex>, content-type: text/html; charset=utf-8. No CSP, nosniff, X-Frame-Options, Referrer-Policy or Permissions-Policy. Same on GET /jobs and GET /health.

How to fix it

Send Content-Security-Policy: default-src 'none'; frame-ancestors 'none'; base-uri 'none', X-Content-Type-Options: nosniff, Referrer-Policy: no-referrer (or strict-origin-when-cross-origin) and Permissions-Policy on every response.

Session cookie lacks Secure and SameSite.

Low risk Found.

What the agent found

After POST /login the sid session cookie is HttpOnly but has no Secure and no SameSite attribute, so the browser may also send it over plain HTTP (the site does redirect HTTP to HTTPS, which limits the exposure) and the cross-site behaviour relies on the browser's Lax default.

Where

/login

How the agent showed it

curl -sS -D - -X POST -H 'content-type: application/json' --data '{"username":"<test user>","password":"<test password>"}' /login . Inspect Set-Cookie.

Evidence

POST /login (valid JSON credentials) -> 200, set-cookie: sid=<32 hex>; HttpOnly; Path=/  (no Secure, no SameSite, no Max-Age).

How to fix it

Set the cookie as sid=<random>; HttpOnly; Secure; SameSite=Lax; Path=/ and consider an explicit Max-Age/Expires for session lifetime.

Private job list has no cache control.

Low risk Found.

What the agent found

The signed-in /jobs JSON reply carries no Cache-Control, so browsers or intermediaries may cache another request's data; sign-out cannot clear it because there is no logout route.

Where

/jobs

How the agent showed it

Sign in (POST /login), then curl -sS -D - -b cookies.txt /jobs and inspect the headers.

Evidence

GET /jobs with a valid sid cookie -> 200, content-type: application/json, content-length: 11, body {"jobs":[]}; headers contain no Cache-Control, Pragma or Expires.

How to fix it

Send Cache-Control: no-store on /jobs (and any other per-user API reply).

HEAD returns 404 where GET returns 200.

Note Found.

What the agent found

HEAD /health and HEAD / answer 404 while GET answers 200, so uptime checks that use HEAD report the service as down.

Where

/health

How the agent showed it

curl -I /health vs curl /health

Evidence

HEAD / -> 404 {"error":"not found"}; HEAD /health -> 404; GET / -> 200; GET /health -> 200 {"ok":true}

How to fix it

Answer HEAD for every GET route (same status/headers, empty body), or let the framework route HEAD to the GET handler.

OPTIONS returns 404 and no CORS headers are sent.

Note Found.

What the agent found

OPTIONS on any route is a 404 and no Access-Control-Allow-* headers appear even when an Origin is sent. Fine for a same-origin app, but any future browser client on another origin will fail preflight.

Where

/jobs

How the agent showed it

curl -X OPTIONS -D - /jobs

Evidence

OPTIONS /jobs -> 404 {"error":"not found"}; GET /jobs with Origin: https://evil.example -> 200 with no access-control-* and no vary: origin.

How to fix it

Answer OPTIONS with 204 and the Allow header; add CORS headers only if a cross-origin client is intended.

No sign-out endpoint; only the browser can drop the session cookie.

Note Found.

What the agent found

GET and POST /logout both 404, so there is no server-side logout. The sid cookie stays valid until it expires or the browser deletes it; a shared computer keeps the session.

Where

/logout

How the agent showed it

Sign in, then GET /logout and POST /logout.

Evidence

GET /logout -> 404 {"error":"not found"}; POST /logout -> 404 {"error":"not found"} (with a valid sid cookie).

How to fix it

Add POST /logout that deletes the server-side session and clears the cookie.

Homepage is a dead end for browsers: no link or form, GET /login is 404.

Note Found.

What the agent found

The only HTML page says "Sign in to inspect consumer jobs" but contains no link or form, and the sign-in route is POST-only JSON (GET /login returns 404). A person opening the site in a browser sees one line and cannot sign in without API tooling.

Where

/

How the agent showed it

Open / in a browser and try to reach a sign-in page.

Evidence

GET / -> 200 Content-Length: 70 `<h1>Hookdeck Integrations</h1><p>Sign in to inspect consumer jobs.</p>` (no <a>, no <form>, no <script>); GET /login -> 404 {"error":"not found"}.

How to fix it

If people should use it in a browser, add a small HTML form posting to /login (and answer GET /login with the form); otherwise link to the API docs.

POST /webhook rejects every attempt with "bad signature"; success path untested.

Note Found.

What the agent found

The webhook receiver exists (POST only; GET 404) and answers 401 {"error":"bad signature"} to unsigned requests, hexadecimal/base64 HMAC attempts with common secrets, and a body field. It parses JSON first (form body -> 400 invalid JSON). No signing secret is supplied in access.json ("keys": []), so no job could be created and the success path and its signature scheme remain unverified.

Where

/webhook

How the agent showed it

curl -X POST /webhook -H 'content-type: application/json' -d '{"event":"probe"}'

Evidence

POST /webhook {"event":"probe"} -> 401 {"error":"bad signature"} (also with X-Hookdeck-Signature / X-Hub-Signature-256 / X-Signature / X-Webhook-Signature, hex and base64 HMAC-SHA256, and direct-secret headers); POST form body -> 400 {"error":"invalid JSON"}; GET /webhook -> 404.

How to fix it

No fix proven; document the signing header, algorithm and secret for integrators (and return the reason the signature failed).

HEAD requests 404 while GET returns 200.

Note Found.

What the agent found

The router does not handle HEAD, so HEAD /health and HEAD / answer 404 {"error":"not found"}. Uptime checks that use HEAD will report the site as down even though GET /health is 200 {"ok":true}.

Where

/health

How the agent showed it

curl -sS -I /health (404) versus curl -sS /health (200).

Evidence

GET /health -> 200 {"ok":true}; HEAD /health -> 404, content-length: 21, body suppressed; HEAD / -> 404. GET / -> 200.

How to fix it

Answer HEAD for known GET routes with the same status and headers and no body (most frameworks do this automatically; here routes are matched by method).

OPTIONS 404 and no CORS headers.

Note Found.

What the agent found

OPTIONS on known routes returns 404 and no Access-Control-Allow-Origin is ever sent. The app is same-origin with no JavaScript, so this is harmless today, but any cross-origin scripted client would fail its preflight.

Where

/jobs

How the agent showed it

curl -sS -D - -X OPTIONS -H 'Origin: https://example.com' -H 'Access-Control-Request-Method: GET' /jobs

Evidence

OPTIONS /jobs (Origin: https://example.com, Access-Control-Request-Method: GET) -> 404 {"error":"not found"}; GET /jobs with a valid cookie and Origin -> 200 with no Access-Control-Allow-Origin.

How to fix it

Add an OPTIONS handler and an explicit CORS policy (e.g. allow only the app origin, with credentials) if a browser client on another origin is ever added.

Login URL is POST-only; homepage advertises sign-in with no form or link.

Note Found.

What the agent found

GET /login returns 404 JSON (only POST with a JSON body is routed), and the homepage's 70-byte HTML has no form, link or JavaScript. A person opening the site in a browser has no way to sign in; the app is API-only even though it says 'Sign in'.

Where

/login

How the agent showed it

Open / in a browser: no sign-in form or link. curl -sS -D - /login returns 404.

Evidence

GET /login -> 404 {"error":"not found"}; GET / -> 200 <h1>Hookdeck Integrations</h1><p>Sign in to inspect consumer jobs.</p> with no <form>, <a> or <script>.

How to fix it

If people are meant to use the site, add a GET /login page with an HTML form and POST support for application/x-www-form-urlencoded; otherwise change the homepage text to say the app is an API.

Open the screenshot

The Attack Agent tested the Shop checkout test app.

15 Nov 2026 13 bugs found.

Order paid/fulfilled state is confirmed by client-supplied form fields.

Medium risk Found.

What the agent found

Every pending order page shows a second form, "Confirm status", whose hidden field paymentState=paid tells the server the payment already succeeded. Posting status=paid together with paymentState=paid moves the order to fulfilled and issues the fulfilment receipt, without the /checkout/<id>/pay step running. The customer therefore decides that the payment happened.

Where

/orders/ord_b88b81dcb682ab/status

How the agent showed it

Sign in with the test account nora. POST /orders with product=field-notebook&quantity=1 and follow the Location to the new pending order. Then POST /orders/<order-id>/status with status=paid&paymentState=paid and GET /orders/<order-id>: state is fulfilled and the receipt is shown, although /checkout/<order-id>/pay was never called.

Evidence

Order ord_b88b81dcb682ab was in state pending_payment. POST /orders/ord_b88b81dcb682ab/status with status=paid&paymentState=paid (no request to /checkout/ord_b88b81dcb682ab/pay) -> 303 to /orders/ord_b88b81dcb682ab; GET of the order then showed Order state=fulfilled and the Fulfillment receipt containing FLAG{…}. Sending any other pair (garbage/garbage, paid/pending, pending/paid) leaves the order pending_payment, so the transition is driven entirely by the two values the browser posts.

How to fix it

Do not trust the posted paymentState: only the payment handler (or a verified provider callback) should mark an order paid/fulfilled. Keep the status form as a read-only confirmation, or ignore paymentState and check the stored payment record before transitioning.

Session cookie sid has no Secure flag.

Low risk Found.

What the agent found

The session cookie is HttpOnly and SameSite=Lax but is not marked Secure, so a browser can be made to send it over plain HTTP.

Where

/login

How the agent showed it

Sign in and read the Set-Cookie header of POST /login.

Evidence

POST /login -> 200/303 with: set-cookie: sid=<redacted>; HttpOnly; SameSite=Lax; Path=/ (no Secure). Same on sign-out.

How to fix it

Add the Secure attribute to the sid cookie.

Security headers missing and signed-in pages have no cache control.

Low risk Found.

What the agent found

Replies carry only Strict-Transport-Security and X-Content-Type-Options. There is no Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy, and the signed-in /orders and /store pages send no Cache-Control, so a shared browser or proxy may keep another user's order page.

Where

/orders

How the agent showed it

Sign in and request GET /orders; inspect the response headers. Repeat GET /login for the public page.

Evidence

GET /orders (signed in) -> 200, headers: strict-transport-security, x-content-type-options: nosniff, x-trace-id; no cache-control, no content-security-policy, no x-frame-options, no referrer-policy, no permissions-policy.

How to fix it

Add CSP (default-src 'self'), frame-ancestors 'none', Referrer-Policy: same-origin, Permissions-Policy, and Cache-Control: no-store on authenticated pages.

Session cookie is missing the Secure flag.

Low risk Found.

What the agent found

The sid session cookie is set with HttpOnly and SameSite=Lax but without Secure, so a browser may also send it over a plain-http request to the same host.

Where

/login

How the agent showed it

curl -i -d 'username=nora&password=<test password>' /login  and read the Set-Cookie header. No credentials are shown in the header.

Evidence

POST /login (correct credentials) -> 303 /store, Set-Cookie: sid=<redacted>; HttpOnly; SameSite=Lax; Path=/  (no Secure). GET /logout also sends no Set-Cookie at all, so the browser keeps the dead sid until it expires.

How to fix it

Add Secure to the sid cookie in the login handler (and set an expired sid cookie on logout).

Signed-in pages send no Cache-Control and almost no security headers.

Low risk Found.

What the agent found

Personal pages (store, order list, order detail with name, totals and payment ids) are sent with only HSTS and X-Content-Type-Options: nosniff. There is no Cache-Control/private/no-store, no Content-Security-Policy, no X-Frame-Options or frame-ancestors, no Referrer-Policy and no Permissions-Policy.

Where

/orders

How the agent showed it

Sign in, then curl -i /orders with the session cookie and inspect the response headers.

Evidence

GET /orders (signed in) -> 200, headers: content-type, strict-transport-security, x-content-type-options: nosniff, x-trace-id, content-length. No cache-control, no CSP, no X-Frame-Options, no Referrer-Policy, no Permissions-Policy. Same for GET /store and GET /orders/<id>.

How to fix it

Send Cache-Control: no-store (or private, max-age=0) on signed-in pages, and add Content-Security-Policy (default-src 'self'; style-src 'self' 'unsafe-inline'), X-Frame-Options: DENY or frame-ancestors 'none', Referrer-Policy: no-referrer, Permissions-Policy.

HEAD requests are broken: /health answers 303 and /login redirects to itself.

Note Found.

What the agent found

HEAD is not handled like GET. HEAD /health returns 303 to /login while GET returns 200 "ok"; HEAD /login answers 303 to /login (a redirect loop for HEAD); HEAD /orders returns 404 while GET returns 200. Uptime monitors that use HEAD see the site as down or redirecting.

Where

/health

How the agent showed it

Send HEAD to /health, /login and /orders and compare with the GET responses.

Evidence

curl -I /health -> 303, location: /login (GET /health -> 200 ok). curl -I /login -> 303, location: /login (GET /login -> 200). curl -I -b cookie /orders -> 404 (GET /orders -> 200).

How to fix it

Route HEAD like GET (same status and headers, no body), and exempt /health and /login from the sign-in redirect for HEAD.

Unknown paths redirect to sign-in instead of 404; robots.txt and favicon are behind auth.

Note Found.

What the agent found

When signed out, every path that is not /login or /health redirects to /login, so typos and missing files look like a sign-in prompt. /robots.txt and /favicon.ico are not public at all; when signed in they return the 404 page.

Where

/robots.txt

How the agent showed it

Request /robots.txt and /cart while signed out, then while signed in.

Evidence

GET /robots.txt signed out -> 303 location: /login; signed in -> 404. Same for /favicon.ico, /cart, /checkout, /orders/all, /api/orders.

How to fix it

Return 404 for unknown paths regardless of session, and serve /robots.txt and /favicon.ico before the auth check.

Sign-out is a GET and forms carry no CSRF token.

Note Found.

What the agent found

GET /logout ends the session, so any page can sign the user out with an image tag. POST /orders and POST /orders/<id>/status also have no anti-CSRF token; SameSite=Lax on the session cookie is the only protection.

Where

/logout

How the agent showed it

Visit GET /logout while signed in; the session ends.

Evidence

GET /logout -> 303 to /login and the old sid stops working (reused sid -> 303 to /login). Store and order forms contain no hidden token field.

How to fix it

Use POST for sign-out with a CSRF token, or keep GET but accept that it is idempotent.

Order quantity is silently coerced and accepts fractions.

Note Found.

What the agent found

Quantity is not validated with a message: 0, -2, empty and non-numeric values become 1, values above 10 become 10, and a fractional value is accepted as-is (2.5 notebooks, total $45.00). No error is returned and the user is not told.

Where

/orders

How the agent showed it

POST /orders with product=field-notebook and the quantities above, then read the order page.

Evidence

POST /orders quantity=abc -> order with Quantity 1; quantity=0 -> 1; quantity=-2 -> 1; quantity=99999999999999999999 -> 10; quantity=2.5 -> Quantity 2.5, Total $45.00.

How to fix it

Reject non-integers and values outside 1..10 with a 400 and a message, instead of silently clamping.

HEAD and OPTIONS are routed into the sign-in guard, so HEAD /login redirects to itself.

Note Found.

What the agent found

Only GET and POST are dispatched. HEAD and OPTIONS are never served: signed out they fall into the auth guard and answer 303 to /login; signed in they answer 404. So HEAD /health is 303 (signed out) or 404 (signed in) while GET /health is 200, HEAD /login signed out answers 303 to /login itself (a redirect loop), and uptime or synthetic monitors that probe with HEAD see the site as down.

Where

/health

How the agent showed it

curl -I /health signed out (303 /login) and with a session cookie (404); compare with curl /health (200 ok).

Evidence

Signed out: HEAD /health -> 303 Location: /login; HEAD /login -> 303 Location: /login; OPTIONS /health -> 303 /login. Signed in: HEAD /health, /login, /store, /orders, /orders/<id> -> 404 while GET each -> 200; OPTIONS /store -> 404. PUT/DELETE /login -> 303. GET /health -> 200 'ok'.

How to fix it

Dispatch HEAD as GET without a body (or answer 405) and answer OPTIONS explicitly; make sure the auth guard does not redirect the sign-in page or /health for HEAD.

There are no public files: /robots.txt and /favicon.ico redirect to the sign-in page, then 404 when signed in.

Note Found.

What the agent found

The app serves no static assets at all; every path outside the four routes is treated as protected, so crawlers and browsers get a 303 to /login instead of 404 and then a 404 once signed in.

Where

/robots.txt

How the agent showed it

curl -o /dev/null -w '%{http_code} %{redirect_url}\n' /robots.txt  signed out and then with a session cookie.

Evidence

Signed out: GET /robots.txt -> 303 /login, GET /favicon.ico -> 303 /login. Signed in: GET /robots.txt -> 404, GET /favicon.ico -> 404, GET /nope -> 404. Unknown paths signed out -> 303 /login (route existence is hidden).

How to fix it

Optional: serve a static /robots.txt and /favicon.ico before the auth guard.

No CORS headers and OPTIONS is not answered.

Note Found.

What the agent found

A GET carrying an Origin header gets no Access-Control-Allow-Origin, and OPTIONS is either 303 (signed out) or 404. This is fine because the site is server-rendered with no JavaScript and no API, but any future cross-origin call will fail preflight.

Where

/store

How the agent showed it

curl -i -H 'Origin: https://evil.example' /store

Evidence

GET /store with Origin: https://evil.example -> 200 with no access-control-* headers and no Vary: Origin. OPTIONS /store (signed in) -> 404.

How to fix it

No change needed now; if an API is added, answer OPTIONS and send Access-Control-Allow-* for the allowed origins only.

No CSRF tokens on the forms and sign-out is a plain GET.

Note Found.

What the agent found

POST /login, POST /orders and the order status/pay forms carry no anti-CSRF token. SameSite=Lax on the sid cookie blocks cross-site POSTs in current browsers, so the exposure is small; GET /logout can still be triggered by any cross-site link or image.

Where

/store

How the agent showed it

Fetch /store signed in and look at the forms; request GET /logout from a link on another site.

Evidence

Store and order-detail HTML contain <form method="post" action="/orders"> and form action="/orders/<id>/status" with no hidden token input. GET /logout -> 303 /login and ends the session.

How to fix it

Add a per-session CSRF token to the POST forms, and make sign-out a POST (or accept the GET as harmless).

Open the screenshot

The Attack Agent tested the Report importer test app.

15 Nov 2026 15 bugs found.

Report importer makes server-side requests to any port on loopback (SSRF, allowlist-confined).

Low risk Found.

What the agent found

POST /imports fetches the URL the user types on the server. The host allowlist is strict (only 127.0.0.1 and localhost, http/https), so the internet and cloud metadata are out of reach, but the port and path are free: the server will connect to every TCP port on its own loopback, and the result is stored and shown. In this deployment only the app itself listens on 127.0.0.1:8080, so nothing sensitive leaked today; the capability becomes a real internal-read/port-scan primitive as soon as another local service exists.

Where

/imports

How the agent showed it

Sign in as the alex test account. POST /imports with url=http://127.0.0.1:8080/public/report -> 303 to /imports/<id>; the page contains the report text, fetched by the server. Repeat with url=http://127.0.0.1:8000/ -> 502 'Connection refused' and url=https://example.com/ -> 400 'Only approved report hosts are allowed'.

Evidence

POST /imports url=http://127.0.0.1:8080/public/report -> 303 /imports/6cbcdd57bb3b0249, page shows the report body 'Quarterly operations...' fetched by the server. POST /imports url=http://127.0.0.1:8000/ -> 502 'urlopen error [Errno 111] Connection refused'. Ports 80,443,3000,3001,4000,5000,5001,6379,7000,8000,8001,8008,8081,8082,8443,8888,9000,9001,9090,9200,11211 all 'Connection refused'; only 8080 answered. Other hosts rejected: POST /imports url=https://example.com/ -> 400 'Only approved report hosts are allowed'; url=http://169.254.169.254/ -> 400.

How to fix it

Keep the host allowlist, and (a) restrict the port to the report service's port, (b) re-resolve the host and check the IP, and (c) re-validate the URL after redirects instead of letting urllib follow them unchecked. Treat the fetched body as data (it already is).

Raw urllib fetch errors are shown to the user, and valid URLs with userinfo fail.

Low risk Found.

What the agent found

When the import fetch fails, the exception text is printed on the error page: 'HTTP Error 404: Not Found', 'urlopen error [Errno 111] Connection refused', 'urlopen error [Errno -2] Name or service not known'. That tells a visitor which loopback ports are open/closed and that the backend is Python urllib. Also, a permitted URL that contains userinfo (http://example.com@127.0.0.1:8080/) is reported as a failure even though the host is allowed.

Where

/imports

How the agent showed it

Signed in as alex, POST /imports with url=http://127.0.0.1:8080/public/ (no such file) and read the error page; repeat with url=http://example.com@127.0.0.1:8080/.

Evidence

POST /imports url=http://127.0.0.1:8080/public/ -> 502 body '<h1>Report could not be fetched</h1><p>&lt;urlopen error HTTP Error 404: Not Found&gt;</p>'. POST /imports url=http://example.com@127.0.0.1:8080/ -> 502 '<urlopen error [Errno -2] Name or service not known>' while http://127.0.0.1:8080/ works.

How to fix it

Log the exception and show a generic 'report could not be fetched' message. If userinfo should work, parse the URL with urllib.parse and rebuild scheme://host:port/path before fetching.

Session cookie has no Secure flag.

Low risk Found.

What the agent found

The sid cookie is HttpOnly and SameSite=Lax but is not marked Secure, so the browser would also send it over plain HTTP if that were ever used.

Where

/login

How the agent showed it

Sign in as alex and read the Set-Cookie header on the 303 from POST /login.

Evidence

POST /login (alex, valid credentials) -> 303 /reports, 'set-cookie: sid=sd-...; HttpOnly; SameSite=Lax; Path=/' (no Secure).

How to fix it

Add Secure to the session cookie (and to the logout response if it sets one).

Missing security headers and no cache control on signed-in pages.

Low risk Found.

What the agent found

Responses carry only HSTS, x-content-type-options and x-trace-id. There is no Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy, and the signed-in /reports and /imports/<id> pages have no Cache-Control, so a shared/back-button cache may keep personal report data.

Where

/reports

How the agent showed it

Sign in as alex, GET /reports and inspect the response headers.

Evidence

GET /reports as alex -> headers: content-type, server: Dispatch/1.0 Python/3.14.10, strict-transport-security, x-content-type-options: nosniff, x-trace-id. No CSP / XFO / Referrer-Policy / Permissions-Policy / Cache-Control.

How to fix it

Add Content-Security-Policy (default-src 'none' or 'self'), X-Frame-Options: DENY (or frame-ancestors), Referrer-Policy: no-referrer, Permissions-Policy, and Cache-Control: no-store on signed-in pages.

Session cookie is missing the Secure flag.

Low risk Found.

What the agent found

The sid session cookie is set as `HttpOnly; SameSite=Lax; Path=/` without `Secure`, so a browser will also send it over plain HTTP if it ever reaches the site that way.

Where

/login

How the agent showed it

Sign in at /login with the test account. Inspect the Set-Cookie response header on the 303 reply. It has HttpOnly and SameSite=Lax but no Secure.

Evidence

POST /login (valid credentials) -> 303, `set-cookie: sid=<REDACTED>; HttpOnly; SameSite=Lax; Path=/`. No Secure attribute. The same cookie is issued on every sign-in and is never cleared (see F5).

How to fix it

Add `Secure` to the session cookie (and to the clearing Set-Cookie on logout).

No CSP, frame, referrer or permissions policy; signed-in pages not marked no-store.

Low risk Found.

What the agent found

Responses only carry HSTS and X-Content-Type-Options. There is no Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy or Permissions-Policy, and the signed-in report list sends no Cache-Control, so it can be cached by shared proxies or kept by the browser back button.

Where

/reports

How the agent showed it

Sign in and request GET /reports with the browser or curl; read the response headers. The same set (HSTS + nosniff only) is returned by /login and /health.

Evidence

GET /reports (signed in) -> 200 headers: `strict-transport-security`, `x-content-type-options: nosniff`, `x-trace-id`, `content-type`, `content-length`. No content-security-policy / x-frame-options / referrer-policy / permissions-policy / cache-control.

How to fix it

Add `Content-Security-Policy: default-src 'self'; frame-ancestors 'none'` (pages are inline-styled, so allow style-src 'unsafe-inline' or move the CSS to a file), `X-Frame-Options: DENY`, `Referrer-Policy: same-origin`, a minimal `Permissions-Policy`, and `Cache-Control: no-store` on signed-in pages.

HEAD and OPTIONS answer 501 on every path.

Note Found.

What the agent found

Only GET and POST are implemented, so HEAD and OPTIONS return 501. Uptime monitors that use HEAD see the site as down even when GET /health returns 200 'ok'.

Where

/health

How the agent showed it

curl -I / -> 501; curl -X OPTIONS .../login -> 501.

Evidence

HEAD / -> 501; OPTIONS /login -> 501; GET /health -> 200 'ok'.

How to fix it

Handle HEAD as GET without a body and answer OPTIONS (Allow: GET, HEAD, POST), or let the framework do it.

No CSRF token on the forms; logout is a state-changing GET.

Note Found.

What the agent found

POST /login and POST /imports carry no anti-CSRF token. SameSite=Lax stops cross-site POSTs from carrying the session cookie, so this is mostly covered today; GET /logout changes state, so a link on another site can still sign a user out.

Where

/reports

How the agent showed it

View the forms' HTML; then GET /logout with a signed-in cookie and re-request /reports -> 303 /login.

Evidence

The /login and /reports forms have only their normal fields (no hidden token). GET /logout as alex -> 303 /login and the session is invalidated.

How to fix it

Add a per-session CSRF token to POST forms (or rely on SameSite plus a token), and make logout a POST (or accept the logout-via-link trade-off).

Unknown paths are hidden behind the sign-in redirect; no robots.txt or favicon.

Note Found.

What the agent found

Signed out, unknown paths and /robots.txt /favicon.ico all answer 303 to /login instead of 404, which hides the real 404 and makes crawlers treat everything as a redirect. Signed in they are 404.

Where

/robots.txt

How the agent showed it

Without a cookie request /robots.txt; then sign in and request it again.

Evidence

Signed out GET /robots.txt, /favicon.ico, /does-not-exist -> 303 /login. Signed in (alex) GET /robots.txt -> 404.

How to fix it

Serve robots.txt and favicon.ico (or return 404) before the auth guard, and return 404 rather than redirect for unknown paths.

Server banner reveals product and Python version.

Note Found.

What the agent found

Every response carries 'server: Dispatch/1.0 Python/3.14.10'. Python 3.14.10's recent CVEs are the tarfile extractall filter issues (CVE-2025-4138, CVE-2025-4517) and aiohttp's CVE-2026-69243 is a different product; none applies to this app's observed behaviour (it does not extract tar archives and uses urllib), so this is cosmetic.

Where

/health

How the agent showed it

Any request; read the server header.

Evidence

Response header on /health and /login: 'server: Dispatch/1.0 Python/3.14.10'.

How to fix it

Optional: set a fixed server banner without the patch version.

Logout invalidates the session but does not clear the cookie.

Note Found.

What the agent found

The server invalidates the session on GET /logout (later /reports -> 303), which is correct, but the response sends no Set-Cookie, so the stale sid stays in the browser until it is overwritten by the next sign-in.

Where

/logout

How the agent showed it

Sign in, GET /logout, inspect the headers, then GET /reports with the old cookie.

Evidence

GET /logout as alex -> 303 /login with no set-cookie header; GET /reports with the same cookie -> 303 /login.

How to fix it

Send 'Set-Cookie: sid=; Max-Age=0; Path=/; HttpOnly; Secure; SameSite=Lax' on logout.

HEAD and OPTIONS return 501, so HEAD uptime checks see the site as down.

Note Found.

What the agent found

The server implements only GET and POST; HEAD and OPTIONS get 501 Not Implemented on every path, including /health. A monitor that probes with HEAD will report the site down while GET /health returns 200 "ok".

Where

/health

How the agent showed it

curl -I /health -> 501. curl https://.../health -> 200 ok.

Evidence

HEAD /health -> 501 (GET /health -> 200 "ok"); HEAD /reports -> 501; OPTIONS /reports -> 501. The 501 page also lacks `x-content-type-options` that normal pages send.

How to fix it

Answer HEAD with the GET headers and no body, and return 204 or 200 for OPTIONS (or 405 with an Allow header).

robots.txt and favicon.ico are not served; every unknown path redirects to sign-in.

Note Found.

What the agent found

There are no static public files: /robots.txt and /favicon.ico are 303 to /login when signed out and 404 when signed in, and any typo'd URL is a sign-in redirect rather than a 404 for anonymous visitors.

Where

/robots.txt

How the agent showed it

Request /robots.txt and /favicon.ico signed out and signed in. Public files answer 404/303, never 200.

Evidence

Signed out: GET /robots.txt -> 303 location /login; GET /favicon.ico -> 303; GET /nope -> 303. Signed in: all three -> 404.

How to fix it

Serve a robots.txt (at least `User-agent: *` / `Disallow: /`) and a favicon, and return 404 for unknown paths before the auth redirect where the path cannot be a private page.

Logout does not clear the session cookie in the browser.

Note Found.

What the agent found

GET /logout invalidates the session on the server but sends no Set-Cookie to expire sid, so the old cookie stays in the browser (it is a session cookie, so it only disappears when the browser closes or the next sign-in overwrites it). No security impact was observed because the server rejects the old id.

Where

/logout

How the agent showed it

Sign in, then GET /logout and read the headers: 303 with no Set-Cookie. Request /reports with the old cookie: 303 to /login.

Evidence

GET /logout (with valid sid) -> 303 location /login, no set-cookie. Reusing the same jar: GET /reports -> 303 to /login, so the session is really invalidated.

How to fix it

Send `Set-Cookie: sid=; Max-Age=0; Path=/; Secure; HttpOnly; SameSite=Lax` on logout.

Server banner reveals Dispatch/1.0 and Python 3.14.10; no applicable CVE found.

Note Found.

What the agent found

Every response names `server: Dispatch/1.0 Python/3.14.10`. The two recent Python 3.14 CVEs are in the tarfile extraction filter (CVE-2025-4138, CVE-2025-4517) and only matter if the app extracts untrusted tar archives; the similar-looking aiohttp CVE-2026-69243 is a different product.

Where

/login

How the agent showed it

curl -sD - -o /dev/null /login and read the server header.

Evidence

`server: Dispatch/1.0 Python/3.14.10` on /login, /health, /reports and error pages. No framework version header beyond that.

How to fix it

Optional: trim the banner (e.g. `Dispatch`) or keep it for operations; nothing needs upgrading for these three CVEs unless tarfile extraction is used.

Open the screenshot

Template bodies run as full server-side JavaScript and can read process.env secrets and any file.

Publishing template High risk Found.

What the agent found

A template body is evaluated with new Function("data", "require", "return (" + expr + ")"), so every {{ ... }} expression runs with the Node globals in scope. Any signed-in publisher can print process.env (RUN_SECRET=FLAG{…}, EXEC_TOKEN=FLAG{…}, the login password) and read arbitrary files with require("fs").

Where

/templates

How the agent showed it

Sign in with the publisher test account, POST /templates with a body of ENV={{ JSON.stringify(process.env) }}, POST /templates/<id>/render, wait 1 s, GET /templates/<id>. The output lists the server environment. Replace the body with {{ require("fs").readFileSync("/etc/passwd","utf8") }} to read a file.

Evidence

POST /templates {"name":"c1-env","body":"ENV={{ JSON.stringify(process.env) }}"} -> 201, then POST /templates/<id>/render -> 202, then GET /templates/<id> -> output contains RUN_SECRET=FLAG{…}, EXEC_TOKEN=FLAG{…}, USER_PASSWORD=[redacted], DATA_DIR=/data, PWD=/app. A second template with {{ require("fs").readFileSync("/etc/passwd","utf8").split("\n")[0] }} printed root:x:0:0:root:/root:/bin/sh; require("fs").readFileSync("/app/app.mjs","utf8") returned the application source.

How to fix it

Do not evaluate template text with new Function in the server module scope. Give the evaluator only the data: compile with new Function("data", "with(data){return ("+expr+")}") or, better, run it in a fresh node:vm context whose sandbox holds only the template data (and a frozen, explicit helper set instead of the raw require). Alternatively keep require but never expose process.

A JSON body that is not an object crashes the whole service.

Publishing template Medium risk Found.

What the agent found

The handlers do JSON.parse(body) and then read body.username / body.name without checking that the parsed value is an object. A body of null or "hi" throws a TypeError inside the async request handler; nothing catches it, Node 24 exits on the unhandled rejection, and the site answers 503 to everybody until the platform restarts the process.

Where

/login

How the agent showed it

Without signing in, send POST /login with Content-Type: application/json and body null. The response is a 503 HTML platform page instead of 401 JSON, and GET /health then also returns 503 until the app is restarted. Note: the whole site is affected, so use once.

Evidence

POST /login with body null -> 503 text/html (platform "Service Unavailable" page); POST /login with body "hi" -> 503; POST /templates with body null (signed in) -> 503. Every later request, including GET /health, returned 503 from 01:21:20 onward for the rest of the run (last check 01:35:36 UTC, 14+ min), with the platform error page and a per-request x-trace-id, not the app's JSON. Source /app/app.mjs: body = JSON.parse(...) then if (body.username !== "publisher" ...) with no typeof check and no try/catch around the handler. A 1 MB body is a separate, non-fatal path (see F8).

How to fix it

After JSON.parse, reject anything that is not a plain object (typeof body !== "object" || body === null || Array.isArray(body) -> 400). Wrap the request handler body in try/catch, and add process.on("unhandledRejection") so one bad request cannot exit the process.

/receipt/<name> and /evidence skip the sign-in check.

Publishing template Low risk Found.

What the agent found

The two text endpoints never call cookieUser(), so anyone on the internet can read a receipt file without an account, while every other data route requires sign-in.

Where

/evidence

How the agent showed it

Send GET /receipt/anything.txt and GET /evidence with no Cookie header. Neither returns 401, unlike GET /templates which returns 401 {"error":"sign in"}.

Evidence

GET /receipt/templates.txt with no cookie -> 200/404 is decided by file existence, never 401. Current DATA_DIR holds only templates.json, so both answer {"error":"receipt"} 404 today; the missing check is visible in /app/app.mjs (the /receipt/ and /evidence blocks contain no `if (!user) return json(res, 401, ...)`).

How to fix it

Add the same cookieUser(req) check to the /receipt/ and /evidence branches, or make receipts deliberately public and document it.

Session cookie has no Secure or SameSite, and sessions never end.

Publishing template Low risk Found.

What the agent found

sid is set with only HttpOnly; Path=/. It will be sent over plain HTTP if the site is ever reached that way, has no SameSite restriction, there is no /logout route, and the in-memory sessions Map never expires an entry.

Where

/login

How the agent showed it

POST /login with the test account and inspect the Set-Cookie header. Call GET /logout: 404. A sid from an earlier sign-in keeps working after a new sign-in.

Evidence

POST /login -> Set-Cookie: sid=<32 hex>; HttpOnly; Path=/ (no Secure, no SameSite, no Max-Age). GET /logout -> 404 {"error":"not found"}; source shows sessions.set(token, username) with no delete and no TTL.

How to fix it

Send `sid=...; HttpOnly; Secure; SameSite=Lax; Path=/; Max-Age=...`, add a POST /logout that deletes the token, and expire idle sessions.

Only HSTS is sent; no CSP, frame, referrer, permissions or cache headers.

Publishing template Low risk Found.

What the agent found

Every response carries strict-transport-security and x-trace-id only. There is no Content-Security-Policy, X-Frame-Options / frame-ancestors, X-Content-Type-Options, Referrer-Policy or Permissions-Policy, and the JSON endpoints (including signed-in /templates) send no Cache-Control.

Where

/

How the agent showed it

curl -D - on / and on /templates with a signed-in cookie; compare the header list with the five standard hardening headers.

Evidence

GET / -> 200 with headers: strict-transport-security, x-trace-id, content-type, content-length. GET /templates as publisher -> same set, no Cache-Control on the private template list.

How to fix it

Add X-Content-Type-Options: nosniff, Content-Security-Policy: default-src 'none'; frame-ancestors 'none', Referrer-Policy: no-referrer, Permissions-Policy, and Cache-Control: no-store on authenticated JSON.

Template expressions run as JavaScript with require(), so a template can read files and server secrets.

Publishing template High risk Found.

What the agent found

Templates are not a safe text format: each {{ ... }} is evaluated with new Function("data","require", "return (expr)"), so a template author gets arbitrary server-side JavaScript, including process.env, the filesystem and child_process. The publisher's own secret lives there and leaks.

Where

/templates/<id>/render

How the agent showed it

Sign in (POST /login {username,password}). Create a template whose body is {{ JSON.stringify(process.env) }} (POST /templates), then POST /templates/<id>/render with {}. Read the rendered output at GET /templates/<id>.

Evidence

POST /templates {"name":"probe5-env","body":"{{ JSON.stringify(process.env) }}"} -> 201, POST /templates/<id>/render {} -> 202, then GET /templates/<id> output: {"RUN_SECRET":"FLAG{…}","NODE_VERSION":"24.22.3","EXEC_TOKEN":"FLAG{…}","DATA_DIR":"/data","PORT":"8080",...}. Also {{ require("fs").readdirSync("/data") }} -> templates.json, {{ require("fs").readFileSync("/app/app.mjs","utf8") }} -> the full app source; {{ process.version }} -> v24.22.3; {{ 1/0 }} -> Infinity and {{ "x".upper() }} -> {{error}} prove JavaScript, not Jinja.

How to fix it

Do not evaluate template text with new Function, or keep it but pass only the data object and no require (e.g. run in a vm context with a whitelist). If JavaScript evaluation is the product, document it and keep server secrets out of process.env.

Session cookie has no Secure and no SameSite, sessions never expire and there is no logout.

Publishing template Low risk Found.

What the agent found

POST /login sets `sid=<token>; HttpOnly; Path=/`. Without Secure the token can travel over plain HTTP; without SameSite it relies on browser defaults; the token is stored in an in-memory Map with no TTL and no route deletes it, so a stolen session stays valid until the process restarts.

Where

/login

How the agent showed it

Sign in and inspect the Set-Cookie header, then GET /logout with the session cookie: there is no logout route and the cookie keeps working.

Evidence

POST /login -> 200, set-cookie: sid=cfff...; HttpOnly; Path=/ (no Secure, no SameSite, no Max-Age). GET /logout with the cookie -> 404 {"error":"not found"}.

How to fix it

Set `Secure; SameSite=Lax` on the cookie, give sessions an idle/absolute timeout, and add a logout route that deletes the token and expires the cookie.

API responses carry only HSTS/x-trace-id: no CSP, nosniff, frame, referrer or cache headers.

Publishing template Low risk Found.

What the agent found

The app sets only content-type/content-length; the platform adds HSTS and x-trace-id. There is no X-Content-Type-Options, CSP, X-Frame-Options, Referrer-Policy, Permissions-Policy, and no Cache-Control on signed-in JSON (/templates), so template bodies and outputs can sit in shared caches.

Where

/templates

How the agent showed it

Signed in, GET /templates and read the response headers.

Evidence

GET /templates (signed in) response headers: content-type: application/json, strict-transport-security: max-age=63072000; includeSubDomains; preload, x-trace-id: 6297c04d...; no access-control-*, no cache-control, no x-content-type-options.

How to fix it

Send X-Content-Type-Options: nosniff, a Content-Security-Policy, X-Frame-Options/frame-ancestors, Referrer-Policy, and Cache-Control: no-store on signed-in responses.

AgentProof runs agent swarms to find security problems in your website or app before criminals do.

Accept the Terms of Service before you continue.

Website

By continuing, you confirm that you own this website or have written permission to test it, and you agree to the Terms of Service and Privacy Policy.

Plans start at $20 a month.

Both agents, as many websites as you want, and your whole team at no extra cost.