Trust & transparency

Live numbers

Verifiable security claims

Mirror of SECURITY.md §3. Each row is a claim plus the command an external auditor runs to check us right now. Items marked rolling out are wired in the deploy pipeline but not yet announced as enforced — SECURITY.md remains the source of truth.

cosign verify \
  --key https://cogos.5ceos.com/cosign.pub \
  cogos5ceos.azurecr.io/cogos-api:vN

curl -isX POST https://cogos.5ceos.com/v1/chat/completions \
  -H "Authorization: Bearer sk-cogos-…" \
  -d '{"model":"cogos-tier-b","messages":[{"role":"user","content":"ping"}]}'
# Re-compute HMAC client-side; recipe at /cookbook#verify-signature

git clone https://github.com/5CEOS-DRA/llm-determinism-bench
cd llm-determinism-bench && cat README.md

# In this repo:
grep -n "createHash\|sha256" src/*.js

# Should 401:
curl -sI https://cogos.5ceos.com/admin/keys

# Should 400 (signature missing):
curl -sI -X POST https://cogos.5ceos.com/stripe/webhook \
  -H "Content-Type: application/json" -d '{}'

# See /cookbook recipe 1 — strict integer schema, prompt for a string answer, get an integer.

Recent revisions

Published security advisories

Per-call receipt chain (always-on)

This is the load-bearing continuous-verification mechanism — it runs on every customer call, not on a schedule. Every /v1/* response carries a receipt with request_id, deterministic_hash, output_hash, and (for grounded chat) verify_verdict. Each usage row is HMAC-chained to the previous row in a per-tenant append-only log; a chain break would be detectable in 90 seconds via GET /v1/audit?since=<ms> where the response includes chain_ok: true/false computed server-side from the returned slice.

The chain head is submitted to OpenTimestamps calendar servers on every audit-checkpoint with new activity (the checkpoint scheduler ticks hourly; idle hours don't submit). Bitcoin confirmation latency from submission to a block including the calendar's aggregate is typically 3-12 hours depending on calendar batching and block timing — so a customer checking the most recent confirmed anchor may see one that's several hours behind the latest submitted checkpoint. The anchored timestamps are publicly verifiable using the open-source ots verify CLI without contacting our infrastructure.

Verify yourself: pick any of your own request IDs, fetch the audit slice, re-hash the row, and compare to row_hash. A published Python re-verifier is on the roadmap; today the check is a one-liner (SHA-256 of the row's canonical bytes, compared to row_hash).

Verify a receipt yourself · in your browser

Paste any x-cogos-attestation header value you've received from a /v1/* response. Your browser fetches our public key from /attestation.pub, then verifies the ed25519 signature client-side using the Web Crypto API. Nothing leaves your browser except that public-key fetch.

Public hash-chain checkpoint

Every hour we snapshot every (tenant_id, app_id) chain head into a globally hash-chained checkpoint. Capture global_head now, replay /audit/checkpoint?ts=<your_capture_ts> later, prove we haven't rewritten any row in between.

Verify yourself:

curl -s https://cogos.5ceos.com/audit/checkpoint/latest
curl -s https://cogos.5ceos.com/audit/checkpoint/verify

Scheduled probe job

The continuous unauthenticated-probe job (scripts/probes-unauth.sh + scripts/pentest-authed.sh in the repo) is currently not scheduled. The previous scheduler was dropped in the 2026-06-20 Azure→AWS cutover and has not yet been re-stood-up as a systemd timer on the new host. The probe scripts themselves are operator-runnable on demand. Re-enabling the schedule is on the roadmap; status will appear here once a run lands.

Pentest summary

Coordinated disclosure