You don't know what's running
Models show up on developer laptops, in vendor SaaS, embedded in features no one filed a memo for. The first risk register is always wrong.
Verifiable AI governance — Apache-2.0, no SaaS lock-in
Find every AI on your network. Sign what matters. Hand the auditor a bundle they can verify themselves.
Shadow AI to signed evidence — in one afternoon. Beacon discovers the models running in your environment, attests them against 23 global audit frameworks — NIST AI RMF, EU AI Act, ISO/IEC 42001, HIPAA, GDPR, Colorado AI Act, Korea AI Basic Act, and more — then produces a cryptographically signed bundle anyone can re-verify with one command.
Live receipts — Ed25519 / JCS canonical
idle
--:--:--awaitingPress “See it sign” to begin the demo—
60-second elevator pitch
Watch Beacon do it.
From shadow AI to signed evidence to a green risk score — in one minute.
The problem
Shadow AI
PDF theater
Aspiration without proof
Three failure modes every AI program hits.
Beacon was built because none of these are solvable with another policy PDF.
Your evidence is a screenshot
Audits today are PDFs of dashboards of claims. Nothing is verifiable months later. Nothing survives a key rotation. Nothing is portable.
“We follow NIST AI RMF” — show me
Frameworks turn into checkboxes with no underlying evidence. Beacon makes the checkbox and the cryptographic receipt the same artifact.
Act 1 of 3 · GET TO YES
Act 2 · Stay at yes
Act 3 · Recover to yes
Ship AI — discovery, not paperwork.
Paste an inventory CSV or let the demo run itself. Beacon tags every model with vendor, capability, residency, and the framework articles it must answer to.
Discovery — Act 1
Act 1 · Ship
Act 2 of 3 · STAY AT YES
Act 3 · Recover
Steady AI — every claim is a signed receipt.
For each model, Beacon serializes the attestation in JCS canonical form (RFC 8785), signs it with an Ed25519 key, and emits an NDJSON event. Click any receipt below to see the exact bytes that were signed — and verify the signature in your browser.
Receipt — JSON view
Canonical bytes (signed)
Ed25519 ✓ verified in-browser
The bargain: agents do the bureaucracy — discovery, classification, evidence collection, signing. Humans hold the moral legitimacy — the policy, the exception, the override. Beacon is the audit log between them.
Act 1 · Ship
Act 2 · Steady
Act 3 of 3 · RECOVER TO YES
Recover AI — hand the auditor a bundle they verify themselves.
When the receipts settle, Beacon assembles them into a portable evidence bundle: the manifest, the NDJSON receipt log, the public keys, the policy YAMLs, the checklists, and a one-page VERIFY.md. Anyone with openssl can re-check it. No vendor login. No phoning home.
Policy as code
Model beacon
Prompt beacon
Artifact beacon
Three beacons watch your network. Privacy by default.
Manual checklist first — wizard walks each control. Then deploy three passive collectors. They never store prompts or outputs. Only signed metadata that proves your AI is governed.
Discovers every model in use
Watches egress to known AI API hosts (OpenAI, Anthropic, Bedrock, Vertex, Azure OpenAI, Cohere, Mistral, Groq, Perplexity, DeepSeek…) and resolves to processes. One signed receipt per (host, process) pair.
No payloads stored. Metadata only.
Counts traffic, not content
Log-tail mode reads existing CASB / SIEM exports (Splunk, Sentinel). Extracts model name, token count, route. Optional TLS proxy mode is opt-in and clearly labelled. Default is privacy-preserving.
Configurable retention. Signed every run.
Tracks AI-generated outputs
Watches S3 buckets, SharePoint folders, Confluence spaces. Detects C2PA provenance markers and AI-generation signatures. Tags every artifact with a receipt that follows it through the org.
Provenance you can verify months later.
See the beacons source on GitHub ↗ · Or run the manual wizard first ↗
Daily risk score
One number your board will actually understand.
Every receipt rolls into a weighted score per framework. Frameworks roll into one AI Risk Index from 0 (no evidence) to 100 (every control attested). Updated daily. Reproducible from the receipts alone.
55.4
Sample AI Risk Index
| Framework | Score | Status |
|---|---|---|
| ISO/IEC 23894 | 80.3 | Strong |
| SOC 2 TSC | 69.9 | Strong |
| ISO/IEC 42001 | 65.8 | Watch |
| NIST AI RMF | 58.1 | Watch |
| EU AI Act | 42.6 | Watch |
| NYC LL 144 | 25.9 | Gap |
| India DPDP | 9.4 | Gap |
The 100-case proof set
Every public AI failure we could find — mapped to the receipt that would have caught it.
From Zillow's $304M write-down to Hong Kong's $25M deepfake wire fraud to Cruise's robotaxi pedestrian drag. Each row links the incident to the framework controls that were missing and the Beacon receipt — model, prompt, artifact, or checklist — that would have produced the evidence. Filter by sector, act, or framework.
100Documented failures
—YES-Ship AI
—YES-Steady AI
—YES-Recover AI
Loading 100 cases…
Sources are primary where possible (regulator filings, court documents, Reuters/AP/NYT/Bloomberg/FT). Methodology and full JSON: ai_failures_top100.json.
The framework registry
23 audit frameworks. Global. In force and in draft.
Every framework ships as machine-readable YAML and XML — same schema, mapped to control IDs from the source text. Auto-translation into any written language is wired in (Spanish and Arabic shipped today). Filter by status and jurisdiction.
Loading framework registry…
The complete audit framework — all 23 lenses
One PDF, every framework, every control row, every evidence prompt. Hand this to your audit committee. Or grab the full YAML/XML registry for your own tooling.
Run it yourself
Three commands. Your own keys. Your own evidence.
No SaaS to sign up for. Beacon is a small Node server plus a static Studio. Apache-2.0 licensed. Designed to run on Render, Railway, or DigitalOcean — or on your laptop.
# 1. Get the code git clone https://github.com/bobrapp/aigovops-beacon.git cd aigovops-beacon # 2. Install & generate your signing key pnpm install pnpm beacon keys:generate # 3. Run discovery → sign → bundle pnpm beacon discover --csv inventory.csv pnpm beacon attest --all pnpm beacon bundle --out evidence.zip # 4. Anyone can verify, no Beacon required pnpm beacon verify evidence.zip # or with stock openssl — VERIFY.md walks you through it
Deploy in one click
Spin up the Studio (Beacon's web UI) on a managed tier. Render and Railway have free starter tiers; DigitalOcean App Platform uses a paid basic instance.
Required: Node 20+, pnpm 9+. No database. Receipts are append-only files.
The Foundation
“Agents do the bureaucracy; humans hold moral legitimacy.”— AIGovOps Foundation, founding principle
Beacon is the reference implementation of an AIGovOps Foundation idea: that AI governance should be boringly verifiable, not theatrically reported. Built and open-sourced by Bob Rapp and Ken Johnston for the AIGovOps Foundation. Apache-2.0 — fork it, sign with your own keys, run it your way.