← Back to home
Security at BidDeed.AI
Last reviewed: October 2026
No fake badges. Everything below is either directly verifiable (a live endpoint, a public repo file, a database query) or explicitly labeled as in-progress. We don't display compliance marks we haven't earned.
🔒 Data Protection
- TLS in transit on every request (Vercel + Cloudflare edge termination).
- Encryption at rest on the primary database (Supabase).
- Row-Level Security is enabled on 723 of 728 public database tables (99%) — verified by direct query against
pg_class.relrowsecurity, October 2026. The remaining tables are reference/lookup data with no customer or credential content. - HTTP security headers (HSTS, CSP, X-Content-Type-Options, X-Frame-Options, Referrer-Policy, Permissions-Policy) on every response from this site — Mozilla HTTP Observatory grade C+, up from F before October 2026. Not yet A: our CSP still allows inline scripts for a PostHog snippet that isn't nonce-based yet — tracked as an open item, not hidden.
🔑 Access Control
- MCP access uses OAuth 2.1 (WorkOS AuthKit) via RFC 9728 protected-resource discovery, or a scoped
bd_live_*API key — both validated server-side on every call. - Secrets are never committed to source. Vault access from any automated agent goes through gated accessor functions with name allow-lists, not direct table reads.
- Cloudflare and GitHub deploy tokens are scoped to the minimum permissions and single zone/repo they operate on.
🤖 AI Security
- Every one of the MCP server's tools runs every call through a single canonical pipeline (auth → cert gate → billing → tool handler) — nothing bypasses it.
- Pattern-based prompt-injection and secret-leak scanning runs on tool arguments and tool results at that pipeline's single chokepoint.
- Ask Deed chat (biddeed.ai/chat): the same kind of pattern-based redaction runs on every message, uploaded document and record block before the model sees it; uploaded documents are fenced as untrusted data, and every redaction is logged. It is a heuristic filter, not a classifier.
- Every tool response carries an explicit notice: scraped county records and case data are untrusted external data, never instructions.
- Ask Deed is reviewed against the OWASP Top 10 for LLM Applications 2026: invisible Unicode characters (zero-width, variation selectors, tag block) are stripped from what goes in and what comes out; answers never load images, and a link to any site other than biddeed.ai shows that site's address; a cached answer is reused only for an identical conversation with identical context, and answers built from your files or a route are never cached. Deed has no tools that write data or move money, and plan or checkout changes in chat need a verified email and an explicit confirm.
- Every tool is billing-gated and idempotency-keyed — a retried or replayed call cannot double-charge or double-execute.
🔍 Audit & Monitoring
- Every pull request to this codebase runs an automated security gate: Semgrep SAST, Gitleaks secret scanning, and dependency audit (npm/pip) — CRITICAL/HIGH findings block the merge. Changes pushed straight to the main branch are covered by the same scan run weekly across the whole codebase (Sundays).
- Error monitoring via PostHog (100K events/month free tier) — unhandled exceptions in the Cloudflare Worker are captured server-side and reported, no browser SDK required.
- A passive OWASP ZAP scan of biddeed.ai and a scripted Ask Deed red-team are scheduled weekly (Mondays); as of October 2026 their first runs are still pending.
- Closed September 2026: legacy vault-reading database functions were callable by the public and signed-in roles; that access was revoked. Re-checked October 2026: 0 of 106 such functions are executable by either role.
📋 Compliance Posture
- SOC 2 Type I — in preparation, not yet certified.
- CASA Tier 2 — planned.
- Florida financial/public-record data handling per FS 197.552 and FS 713.07.
- Governing law: State of Florida. See our Terms of Service.
📚 Security Documentation
- 📄 Incident Response Plan (request a copy) — severity classification, detection sources, response playbooks.
- 📋 Vendor & Sub-Processor List (request a copy) — every third party that touches customer data, with sourced security-page links.
- 🗓 Data Retention & Deletion Policy — what we keep, for how long, and how to request deletion.
- 📊 External Scan Results (request a copy) — Mozilla HTTP Observatory (C+, up from F) + SSL Labs (A on TLS), raw output attached.
- ✅ CAIQ Self-Assessment (request a copy) — CSA-domain-structured control answers, including the gaps, not just the passes.
- 🤖 AI Security Self-Assessment (request a copy) — prompt-injection controls, model governance, and disclosed AI-specific gaps.
- ❓ Common Questionnaire Answers (request a copy) — pre-written answers to the 50 questions vendor security reviews ask most.
- 📦 Request the full Security Evidence Pack (architecture, controls, compliance posture) — email security@biddeed.ai.
🏢 Enterprise Trust Portal
A gated self-serve portal (trust.biddeed.ai, via SafeBase) is in setup — not live yet. Once it launches it will host the documents linked above plus any completed penetration test report, in one request-access location for procurement teams. Until then, email security@biddeed.ai for any of the documents above — same documents, no waiting on the portal.
📧 Security Contact
Found an issue? Email security@biddeed.ai. We aim to respond within 48 hours. Responsible disclosure is welcome — please give us a reasonable window to fix before public disclosure.