EU AI Act · Regulation 2024/1689
The flight recorder for your AI stack.
Open-source scanners that log, check, and prove EU AI Act compliance — before December 2027 makes it mandatory.
$ air scan ./pipeline
✓ art. 12 — logging & traceability
✓ art. 14 — human oversight
✓ art. 15 — accuracy & robustness
✗ art. 10 — data governance (2 findings)
audit chain signed · HMAC-SHA256
recording
8
packages on PyPI
Dec 2027
high-risk obligations apply
100% open
Apache-2.0, no telemetry
Scan
Point it at a repo, a pipeline, or an agent framework. Every check maps to an article, not a vibe.
Record
A tamper-evident audit chain logs what your system did and when. Signed, verifiable, boring — the way evidence should be.
Prove
Export evidence bundles an auditor can actually read — mapped to ISO 42001 and NIST AI RMF.
The Problem
Your AI is making decisions.
Can you prove what it said?
Every company using LLM APIs has the same blind spot. Calls go out. Responses come back. Nothing in between is recorded, verifiable, or auditable.
No audit trail
When a regulator asks "what did your AI tell that customer?", you're digging through logs that don't exist. OpenAI doesn't store your prompts for you.
Plaintext API keys in transit
Every developer on your team has the API key. Every request carries it in cleartext. One leaked key = unlimited spend on your dime.
Zero replay capability
Model got updated? Output changed? Good luck proving what the old version said. Without recorded request/response pairs, you can't reproduce anything.
How It Works
One line change. Complete coverage.
Point your app at the gateway instead of the provider. That's it. Everything else is automatic.
01
Your app sends a request
Change your base URL from api.openai.com to localhost:8080. Same SDK. Same code. Same everything.
02
Gateway authenticates & proxies
The gateway checks your gateway key, forwards the request upstream, and streams the response back in real time. Sub-millisecond overhead.
03
AIR record created
A tamper-evident .air.json record is written asynchronously: request, response, model, tokens, timestamp, run ID. Never blocks your response.
04
Secrets vault-encrypted
API keys and auth headers are stripped from the AIR record and encrypted separately. Even with the audit file, credentials can't be extracted.
Integration
Change one line. That's the integration.
# Before (direct to OpenAI) client = OpenAI( api_key="sk-...", base_url="https://api.openai.com/v1" ) # After (through AIR Gateway) client = OpenAI( api_key="sk-...", base_url="http://localhost:8080/v1", default_headers={"X-Gateway-Key": "your-gateway-key"} )
Features
Built for production. Not a weekend hack.
Security
Gateway-level auth
X-Gateway-Key header authentication. Your upstream API keys never leave the server. Developers hit the gateway, not the provider.
Security
Encrypted vault
Secrets are AES-encrypted and stored in a separate vault (local or S3-compatible). AIR records contain zero plaintext credentials.
Performance
Non-blocking writes
Audit records write in background goroutines. Vault writes are async. Your response latency is the provider's latency. Period.
Performance
SSE streaming
Full support for streaming responses. Tokens stream to your app in real time while the gateway records the complete response.
Compliance
Tamper-evident records
Each AIR record includes cryptographic hashes. If anyone modifies a record after the fact, the hash breaks. Provable integrity.
Compliance
Replay & diff
Replay recorded requests against current models. Compare outputs. Detect when a model update changed behavior. Regression testing for AI.
Trust Layer
Cryptographic audit chain
Every AIR record is linked via HMAC-SHA256 into a tamper-proof chain. Modify any record and the chain breaks. Blockchain-grade integrity without the blockchain.
Trust Layer
SOC 2 & ISO 27001 mapping
Automatic compliance evaluation against 22 controls across SOC 2 Trust Service Criteria and ISO 27001 Annex A, from your live configuration.
Trust Layer
Evidence export for regulators
One API call generates a signed evidence package: audit chain, compliance report, and HMAC attestation. Hand it to your auditor as-is.
Developer Experience
Docker Compose ready
One command to run. No MinIO dependency required. Works with local filesystem or S3-compatible storage. Your choice.
Developer Experience
Provider agnostic
Works with any OpenAI-compatible API. OpenAI, Anthropic (via proxy), Azure OpenAI, local models, custom endpoints. Same format.
Developer Experience
MCP server included
pip install air-blackbox-mcp puts the scanner inside Claude Desktop, Claude Code, and Cursor — scan and remediate from your editor.
Who It's For
If you're calling an LLM API,
you probably need this.
Healthcare
AI-assisted clinical decisions
When your AI suggests a diagnosis or treatment path, regulators will want to see exactly what was asked and what was returned. AIR gives you that record, encrypted and timestamped.
Financial Services
Algorithmic compliance
Trading desks and advisory platforms using LLMs for research or recommendations need audit trails that prove what the model said — not what someone remembers it said.
Legal
AI-generated documents
Law firms using AI for contract review, brief drafting, or legal research need to prove the provenance of AI-assisted work. AIR records create that chain of custody.
Enterprise / SaaS
Customer-facing AI features
Building a chatbot? An AI assistant? When a customer disputes what your AI told them, you need receipts. Every conversation, every token, every timestamp.
By The Numbers
Lightweight. Auditable. Open.
39
automated compliance checks (26 static + 13 runtime)
6
EU AI Act articles covered (9, 10, 11, 12, 14, 15)
8
PyPI packages in the ecosystem
5
framework trust layers (LangChain, CrewAI, AutoGen, OpenAI, RAG)
See It In Action
Four commands. Full coverage.
Run this on any Python AI project and get a compliance report, shadow AI scan, replayable audit trail, and signed evidence package — in under 60 seconds.
air-blackbox comply
Scans your project against 6 EU AI Act articles. Flags HIGH and MEDIUM severity gaps with exact file references.
air-blackbox discover
Finds undeclared model calls hiding in utility functions and helpers — your shadow AI inventory.
air-blackbox replay
Loads any past episode from the audit chain, verifies the HMAC-SHA256 signature, and replays every step with timestamps.
air-blackbox export
Generates a cryptographically signed evidence package with AI-BOM, compliance report, and audit chain — ready for regulators.
Still on when everything else is dark.
When the auditor calls, the question isn't whether your AI worked. It's whether you can prove what it did.
Frequently Asked Questions
Everything you need to know about AIR Blackbox
What is AIR Blackbox?
AIR Blackbox is an open-source EU AI Act compliance scanner for Python AI agents. It runs 39 automated checks across 6 EU AI Act articles (9, 10, 11, 12, 14, 15) against your source code and runtime configuration. It supports LangChain, CrewAI, AutoGen, OpenAI Agents SDK, and RAG pipelines. Install with pip install air-blackbox and scan with air-blackbox comply --scan . -v. The entire tool runs locally — your code never leaves your machine.
What EU AI Act articles does it check?
AIR Blackbox checks 6 articles: Article 9 (Risk Management — error handling, fallbacks, risk assessment), Article 10 (Data Governance — input validation, PII handling, data quality), Article 11 (Technical Documentation — docstrings, type hints, README, model cards), Article 12 (Record-Keeping — logging, tracing, audit trails, observability), Article 14 (Human Oversight — HITL gates, kill switches, rate limits, identity binding), and Article 15 (Accuracy & Security — injection defense, output validation, retry logic). Each check is classified as static (verifiable from source code) or runtime (requires gateway/trust layer).
How does two-tier scoring work?
Two-tier scoring separates the 39 checks into two categories: 26 Static checks analyze code patterns, documentation, and configuration that can be verified from source code alone. 13 Runtime checks require a running gateway or trust layer to verify (PII detection in live prompts, tamper-evident audit chains, kill switches). This gives teams a realistic compliance score even without the full gateway deployed — you can pass all static checks immediately and work toward runtime compliance incrementally.
When is the EU AI Act deadline?
The key deadline for high-risk AI systems is December 2027, pushed back from the original August 2, 2026 date. Penalties for non-compliance can reach up to €35 million or 7% of global annual turnover, whichever is higher. Prohibited AI practices enforcement began February 2, 2025. GPAI model obligations apply from August 2, 2025. The delay moves the audit date, not the evidence problem: records have to exist from the day your system runs, not from the day enforcement starts.
What frameworks does AIR Blackbox support?
AIR Blackbox has trust layer integrations for LangChain (pip install air-langchain-trust), CrewAI (pip install air-crewai-trust), AutoGen (pip install air-autogen-trust), OpenAI Agents SDK (pip install air-openai-trust), and RAG pipelines (pip install air-rag-trust). The compliance scanner works on any Python AI code regardless of framework. There's also an MCP server (pip install air-blackbox-mcp) for Claude Desktop and Cursor integration.
How does it compare to Credo AI, Holistic AI, or OneTrust?
Enterprise AI governance platforms (Credo AI, Holistic AI, Vanta, OneTrust) typically cost $50,000+/year and require sending code to their cloud. AIR Blackbox is free, open source (Apache 2.0), and runs 100% locally. It focuses specifically on EU AI Act technical requirements for Python AI agents. The developer experience is fundamentally different: pip install and scan in 10 seconds, versus weeks of procurement and enterprise deployment. AIR Blackbox is a linter for AI governance — lightweight, fast, and integrated into developer workflows via CLI, pre-commit hooks, and GitHub Actions.
Is AIR Blackbox free?
Yes, AIR Blackbox is 100% free and open source under the Apache 2.0 license. All 8 PyPI packages (scanner, 5 trust layers, MCP server, GitHub Action) are free. There is no paid tier, no freemium model, and no telemetry. The project runs entirely locally by design.
What is the HMAC-SHA256 audit chain?
Every AI action logged through the AIR gateway or trust layers is written as a tamper-evident .air.json record. Each record is linked to the previous one via HMAC-SHA256 cryptographic hashes — creating a blockchain-style chain without the blockchain. If anyone modifies a record after the fact, the hash chain breaks and the tampering is detectable. This provides auditor-grade evidence that AI decisions were recorded accurately and haven't been altered.
EU AI Act high-risk deadline is December 2027.
39 checks. 6 articles. 5 frameworks. One pip install.
$ pip install air-blackbox
click to copy