AIR Blackbox is an open-source AI agent governance framework released in February 2026 by the AIR Blackbox Team, which defines itself as a "black box flight recorder for autonomous AI agents." Just as an aircraft's black box records all flight data in a tamper-proof manner, this tool seals the entire LLM invocation with an HMAC-SHA256 chain and an ML-DSA-65 (FIPS 204, Dilithium3 family) quantum-resistant signature, creating an audit chain where any tampering attempt is immediately detected. Its core operation involves simply replacing the base_url of existing OpenAI/Anthropic API calls with a local gateway, transparently intercepting and recording the entire prompt, response, and tool invocation with approximately 0.3ms (median) overhead per call, without requiring code refactoring. Existing logging systems have three fundamental limitations. First, if someone has write access to the log repository, they can modify the records, making tamper-proof verification impossible. Second, most logging only captures the response and does not fully record the entire prompt context, intermediate reasoning process, and tool invocation chain. Third, raw logs alone cannot satisfy the requirements of Article 12 of the EU AI Act, which mandates tamper-proof logs, retention periods, and guaranteed audit access. AIR Blackbox addresses these three limitations with a four-layer architecture (Verify, Filter, Stabilize, Protect). The Verify layer prevents tampering with HMAC chains and quantum-resistant signatures, the Filter layer blocks 20 patterns of PII detection and prompt injection attacks before they reach the model, the Stabilize layer detects drift in the compliance status of the CI/CD pipeline, and the Protect layer cryptographically guarantees human oversight attestation. While tools like Arthur AI and Lakera focus on a single aspect, input filtering, AIR Blackbox covers the entire lifecycle, from pre-policy enforcement to post-audit evidence packaging. Practical use case scenarios include: First, preparing for EU AI Act compliance audits. With a single line, air-blackbox comply --scan . -v, it runs over 51 static checks against Articles 9-15 for the entire Python AI project and automatically generates reports mapping to four frameworks: ISO 42001, NIST AI RMF, and Colorado SB 24-205. Second, using the Covenant YAML policy in air-gate to pre-define the actions that an agent can perform and issue Ed25519-signed bilateral receipts for all actions, allowing the entire multi-agent delegation chain to be tracked. Third, packaging the audit chain, scan results, SHA-256 manifest, and ML-DSA-65 signature into a self-verifiable .air-evidence ZIP file with the air-blackbox export command, allowing auditors to complete a PASS/FAIL determination in under 2 seconds with a single line, python verify.py, without requiring separate installation. It provides a drop-in trust layer for seven major frameworks, including LangChain, CrewAI, OpenAI Agents SDK, Anthropic Claude SDK, Google ADK, AutoGen, and Haystack, allowing integration with minimal changes to existing agent code.

Open-source AI agent governance and auditing framework.

Pre-Audit for EU AI Act Compliance: Scan the entire Python AI project for 51 items based on Articles 9-15 using air-blackbox comply --scan . -v. Automatically generate mapping reports for ISO 42001, NIST AI RMF, and Colorado SB 24-205. Integrate with CI/CD using GitHub Action (compliance-action) to automatically verify compliance status for each PR.