Report security issues to the maintainers via a GitHub security advisory on this repository ("Report a vulnerability"). Do not open public issues for exploitable bugs. Please include reproduction steps and affected areas. You can expect an initial response within 7 days.

• Authentication: scrypt password hashing (N=16384, per-user salt), httpOnly + SameSite=Lax + Secure-in-production session cookies, 12 h idle / 7 d absolute session timeouts, session list with revoke, session invalidation on password change.
• Authorization: centralized requireProjectAccess / requireOwnedRun / requireOwnedIncident; resources belonging to other users return 404, never 403. Server actions and API handlers check ownership server-side.
• CSRF: mutating server actions verify the request Origin against the request host (browsers always send Origin on mutations), with Sec-Fetch-Site as a fallback; session cookies are SameSite=Lax, blocking cross-site form forgery.
• Rate limiting: sliding-window limits on login, registration, reset, webhook deliveries, run triggers, verification, and AI investigations.
• Webhooks: HMAC-SHA256 over the raw body, constant-time compared before JSON parsing; delivery-ID deduplication with a persisted delivery log; size and content-type checks; 1 MB body cap.
• Secrets: integration credentials encrypted with AES-256-GCM (key from ENCRYPTION_KEY, versioned for rotation); write-only in the UI (last four characters at most); central redact() applied to stored run output and error paths; secrets are never sent to the browser.
• SSRF protection for monitoring: http/https only, credential-free URLs, port allowlist (80, 443, 8080, 8443), DNS resolved and validated against loopback/private/CGNAT/link-local/multicast/reserved ranges (including IPv4-mapped IPv6 and decimal/octal/hex encodings), connection pinned to the validated IP, redirects revalidated (max 4), response-body cap, no cookies forwarded.
• Run binding integrity: runs are created once with their commit SHA and are never updated to a different SHA — enforced in application code and, on PostgreSQL (ADR-010), by a database trigger (test_run_commit_binding) that rejects any UPDATE changing repoOwner/repoName/commitSha; statuses are published only after re-reading the PR's current head and only against the run's own SHA.
• Security headers: CSP (no inline scripts in production), X-Content-Type-Options, Referrer-Policy, Permissions-Policy, frame-ancestors 'none', HSTS in production.

• Dependency advisories (as of build date): npm audit --omit=dev reports 0 vulnerabilities (Next.js 16.4, React 19, Prisma 6.12). The dev-only audit shows a brace-glob stack-exhaustion advisory reachable only from the Tailwind 3 build toolchain (not shipped to production); closing it requires a Tailwind 4 migration, tracked as follow-up work.
• The rate limiter is database-backed (a shared RateLimitWindow table), so limits hold across web instances. Windows are fixed-length; bursts at a window boundary can admit up to 2x the limit briefly.
• Email delivery (notifications and password-reset links) uses a provider HTTP API (Brevo by default, Resend optional via EMAIL_PROVIDER). Without EMAIL_PROVIDER_API_KEY/EMAIL_PROVIDER_FROM the channel reports not_configured honestly: reset links are never displayed or logged, and an operator must retrieve tokens from the database. Configure a provider before relying on email flows.
• There is no account enumeration protection issue: reset/login responses are generic, but login rate limiting is keyed partly by email, which leaks nothing beyond timing.
• Monitoring checks run from the worker process; network-layer egress restrictions (firewalling the worker) are recommended defense-in-depth and are not configurable from the app.
• No CSP nonces are used; the policy is origin-based. Inline <script> is disallowed, which is verified by the build (Next injects no inline scripts in production).

DevSentinel AI is a hackathon project. It implements real defenses, but no software is absolutely secure. Do not deploy it with production secrets without reviewing the limitations above.