Skip to content

Security assessment of the browser session API

Scope: the API a browser uses with a session token, /v1/session/* (session info, conversation, replay, chat over SSE or JSON, cancel, feedback, kiosk exit, bug reports, server voice offer), and what it relies on: token minting (POST /v1/session-tokens), the authenticator, origin checks, the middleware stack and the error handlers. Reviewed on 2026-11-15 by reading the code (api/session_routes.py, api/voice_routes.py, api/deps.py, identity/auth.py, core/security.py, sessions/*, main.py) and testing each point; the fixes are in the same change as this document and covered by tests/test_hardening.py, tests/test_privacy.py and tests/test_guard_*.

Severity: High (exploitable to reach another user's data or the service), Medium (weakens a control or enables abuse), Low (defence in depth, hygiene).

Findings

ID Area Finding Severity Status
S1 Headers API responses carried no security headers: no nosniff, no anti-framing, no Referrer-Policy, no HSTS; the console had a CSP but no HSTS. Low Fixed: api.hardening adds X-Content-Type-Options, X-Frame-Options: DENY, Referrer-Policy (no-referrer for the API, same-origin for the console), Cross-Origin-Opener-Policy, a CSP default-src 'none'; frame-ancestors 'none' on API responses, HSTS when the request came over HTTPS (or X-Forwarded-Proto: https).
S2 Caching Responses holding a session token (POST /v1/session-tokens), a sign-in code or conversations had no Cache-Control, so an intermediary or the browser could keep them. Medium Fixed: Cache-Control: no-store on every API response that sets none (SSE already had it).
S3 Cancel POST /v1/session/turns/{id}/cancel checked only the subject: a token of one conversation (or of another application with the same subject id) could cancel another conversation's turn if it knew the turn id (a random UUID, so not guessable, but disclosed in X-Turn-Id and events). Low Fixed: a token cancels turns of its own conversation only.
S4 Revocation A session token could not be revoked before it expired (up to 15 minutes): no sign-out everywhere, and an erased person's tokens kept working. Medium Fixed: POST /v1/privacy/revoke-sessions and every erase revoke the subject's tokens minted until then (a keyed hash of the subject, a millisecond issue time in tokens, checked on every request with a 5 s cache).
S5 Rate limits No limit on requests per token: a stolen or scripted token could hammer replay, conversation reads and bug reports; turns had only the spending caps. Medium Fixed: 240 requests per token per minute on the whole session API (429 rate_limited), 20 bug reports per user per hour, and the security layer's per-user turn rate, concurrency and conversation length (monitor by default, see security).
S6 Body size No request body limit: a large JSON body (an external history, a page text, a bug report context) was read whole into memory before validation. Medium Fixed: 413 above 2 MB on the session API and 8 MB elsewhere, by Content-Length and by counting streamed bodies.
S7 CORS No CORS handling: cross-origin browsers could not call the API at all (safe, but preflights got 405), so embedding Evokai on another origin needed a same-origin proxy. Low Fixed: preflights and responses of /v1/session/* allow the origins the applications declare, with Authorization and no credentials; the per-application origin check still runs on every request; other paths never answer CORS.
S8 Origin check Requests without Origin are accepted (non-browser clients); the token is a bearer secret. A token stolen from a browser can be replayed from a script until it expires. Low Accepted, mitigated: tokens live 5 minutes by default (at most 15), are bound to one subject, agent and conversation, cannot be minted by a browser, and can now be revoked (S4) and rate-limited (S5). Binding tokens to a client key (DPoP) is a possible next step.
S9 Tokens JWT validation: HS256 only (no none, no algorithm confusion), issuer checked, exp, iat, jti, sub, aud required; console cookies use another issuer so neither passes for the other. aud is the application id and is read rather than verified (the issuer and the key are what matter). Info No change.
S10 Token in URLs Tokens are only sent in Authorization; SSE runs over fetch with a POST, the WebRTC offer too. No token appears in a URL, a log line or a Referer. Info No change.
S11 SSE The stream sends a comment every 10 s, stops at done, cancels the turn when the client goes away and has X-Accel-Buffering: no. Events carry turnId and seq; replay returns only the token's conversation (at most 2,000 events per call). Info No change.
S12 Error leakage Errors are {error, message} with stable codes; unhandled exceptions answer internal without detail; validation errors list field paths only; model and tool failures reach the stream as generic messages. Log lines carry error classes, and now have personal data and secrets masked (pii.logs). Info No change.
S13 Kiosk exit The exit code is hashed with scrypt and limited to 5 attempts per session. Info No change.
S14 Data in bug reports Bug reports store free text and a client context (up to 20 kB) with no masking. Low Partly addressed: erased with the subject, rate-limited (S5); masking free text is out of scope (no reliable detector for names).
S15 Console login throttling Failed console logins are counted per client address; behind a proxy that address is the proxy's unless uvicorn trusts its X-Forwarded-For (FORWARDED_ALLOW_IPS), so one attacker can lock everyone out for 5 minutes. Low Open, documented: the console is internal (SSM port-forward); set FORWARDED_ALLOW_IPS to the load balancer's addresses if it is ever published.
S16 Voice catalogue PUT /v1/admin/voices/{provider}/{id} changes the deployment-wide voice catalogue with any application's admin key. Low Narrowed: console logins need the owner role for it (console roles); bearer admin keys keep their behaviour for compatibility.

Residual risks and recommendations

  • Prompt injection is mitigated, not solved: keep tools least-privilege (the user's own credential, read-only keys for server runs) and confirmations on every change; move input.injection and untrusted.injection to enforce once the security events show no false positives.
  • Per-instance counters (S5 and the security layer's limits) multiply with the number of tasks; a shared store (Redis, the database) would make them global.
  • The deployment's proxy should also enforce TLS and its own rate limits; Evokai's are a second line.
  • Keep the console off the public internet even with SSO and roles.