If you believe you have found a security issue in SecureContent AI, please do not open a public GitHub issue. Instead:

1. Contact the project maintainers privately with:
   • a description of the issue
   • steps to reproduce or a proof-of-concept
   • the commit or file involved
2. You will receive a response within 3 business days.
3. If the issue is confirmed, a fix will be prioritized and a coordinated disclosure date agreed.

This repository ships with layered security controls and CI enforcement (gitleaks, Semgrep, dependency audit, Trivy, secret-in-build checks), but it remains a community-built product: review the hardening checklist before exposing it beyond a trusted operator group.

• Do not upload real credentials, government IDs, or customer data to a shared or public instance.
• All benchmark documents are synthetic — the detectors are heuristic (regex + entropy + context) and do not guarantee complete PII/secret coverage.
• The LLM adapter treats document content as untrusted data inside an explicit <UNTRUSTED_DOCUMENT> envelope and never merges it with system instructions.

• Zero-trust ingestion: every file is validated for size/MIME, parsed in isolation, and scanned before reaching the model.
• Detectors: multi-signal PII / financial / secret / prompt-injection scans with confidence scores and per-document scan options (src/lib/security/detectors.ts). Output DLP always scans everything.
• Policy engine: allow / mask / remove / block buckets over 5 immutable built-in profiles plus user-created custom policies and framework templates (src/lib/security/policies.ts, policy-templates.ts). Credentials and injections can never be allow-listed; injections are always quarantined.
• Sanitization: mask/redact/replace/quarantine plus reviewer per-finding overrides that cannot weaken the invariants (src/lib/security/sanitize.ts).
• Output DLP: deterministic re-scan before release (src/lib/security/output-dlp.ts) plus HTML/URL sanitization (outputSanitizer.ts, incl. javascript:/data:/event-handler stripping).
• API auth (opt-in): set API_AUTH_TOKEN (≥16 chars) and every POST/PUT/DELETE under /api/v1/* requires Authorization: Bearer. Browser token entry lives in the topbar (localStorage only). Set REQUIRE_AUTH_FOR_READS=true to lock GETs too.
• Seed gating: /api/v1/seed runs only in dev/test unless ALLOW_SEED=true (and honors API auth).
• SSRF guard: DOCLING_WORKER_URL is restricted to loopback / RFC1918 / compose service names; metadata (169.254.169.254), public IPs, and credentialed URLs are rejected (isAllowedDoclingUrl).
• Abuse guards: per-IP + global rate limits on expensive routes (bulk/pipeline/seed), strict Zod validation on every trust boundary, security headers (CSP/HSTS/X-Frame-Options/COOP/CORP) in next.config.ts + proxy.ts + Caddy.
• Supply chain: pinned lockfile, overrides for patched transitives, bun audit critical-gate + Trivy container scan in CI.
• Trust boundaries T0–T7 each with validation, auth, logging, and rate-limit notes (see docs/threat-model.md).

• bun audit, gitleaks, semgrep (OWASP + JS/TS + secrets) in .github/workflows/security.yml.
• eslint + tsc --noEmit + vitest in .github/workflows/ci.yml.
• Dependencies are pinned via bun.lock; review bun audit output before merging.

• LLM provider is now server-only Gemini (primary, GEMINI_API_KEY) + Groq backup (GROQ_API_KEY, openai/gpt-oss-120b) via src/lib/ai/transform.ts (runtime server-only guard; keys never leave the server). Rotate keys via env/secrets manager; never expose via NEXT_PUBLIC_*.
• Next.js ≥16.3.5 (fixes GHSA-p293-qw3h-jr36 / GHSA-2xp9-vwfh-vxw4 RCEs + DoS/SSRF advisories), sharp ≥0.35.4, postcss/prismjs overrides; bun audit critical-gate + Trivy in CI.
• Optional bearer auth (API_AUTH_TOKEN), seed gating (ALLOW_SEED), Docling SSRF allowlist, per-IP + global rate limits, hardened compose (read-only FS, no-new-privileges, dropped caps, resource limits, pinned images).
• Put the Docling worker (if used) behind auth and a sandbox; do not expose DOCLING_WORKER_URL publicly.
• Add persistence-layer encryption at rest (SQLite SQLCipher or Postgres + pgcrypto) and TLS everywhere.
• Add RBAC, tenant isolation, and ACL-aware RAG (not just shared SQLite).
• Add WAF, rate limiting, and object-storage scanning for malware.
• Never log rawContent or full findings verbatim — audit logs should store hashes/summaries.
• Run regular dependency scans (pip-audit, npm audit, trivy).
• See docs/cloud.md for the cloud-hosting runbook (TLS, secrets, Postgres, backups, updates).