# AIR Blackbox

> The trust infrastructure between human intent and AI execution. Verify. Filter. Stabilize. Protect.

AIR Blackbox is open-source trust infrastructure that sits inside every AI call -- between your team and the AI stack -- providing decision traceability, escalation intelligence, operational drift detection, and human oversight proof. v1.10.0. 1,500+ tests. 74% coverage. Runs locally. Apache 2.0.

AI made generation abundant. What becomes valuable now is the infrastructure that verifies, routes, constrains, and records machine-assisted work in real time. Compliance is the wedge. Trust infrastructure is the platform.

## Why AIR Blackbox

Enterprise governance platforms (Credo AI, Holistic AI, OneTrust) audit after the fact. AI security firewalls (Arthur AI, Lasso, Lakera) filter threats. AIR Blackbox sits inside the call -- at the interception layer between human intent and AI execution. That gives you:

- **Verify** -- HMAC-SHA256 tamper-evident audit chains for decision traceability
- **Filter** -- PII detection and prompt injection scanning in real time
- **Stabilize** -- 48 compliance checks in CI/CD for operational drift detection
- **Protect** -- Human oversight attestation (Art. 14 delegation logging)

## Key Links

- [Homepage](https://airblackbox.ai)
- [GitHub](https://github.com/airblackbox/gateway)
- [PyPI](https://pypi.org/project/air-blackbox/)
- [Audit Chain Specification (Open Standard v1.0.0)](https://airblackbox.ai/spec)
- [CI/CD Integration Guide](https://airblackbox.ai/ci-cd)
- [Compliance Mapping](https://airblackbox.ai/compliance-mapping)
- [Blog](https://airblackbox.ai/blog/)

## Installation

pip install air-blackbox

## Quick Start

air-blackbox comply --scan . -v

## Features

- 48 EU AI Act + GDPR compliance checks (Articles 9-12, 14-15)
- Article 12 Compliance Layer -- static + runtime analysis for tamper-evident logging
- ML-DSA-65 (FIPS 204) quantum-safe digital signatures
- HMAC-SHA256 tamper-evident audit chains (open standard)
- Self-verifying .air-evidence bundles for auditors
- Prompt injection detection (20 patterns, 5 categories)
- GDPR scanner (8 checks)
- Bias and fairness scanner (6 checks)
- ISO 42001 + NIST AI RMF + Colorado SB 24-205 crosswalk
- Trust layers for LangChain, CrewAI, OpenAI SDK, Claude Agent SDK, Google ADK, AutoGen, Haystack
- Agent-to-agent (A2A) compliance protocol
- Pre-commit hooks
- MCP server integration
- Runtime validation engine (tool allowlists, content policy, PII output scanning)

## Quality

- 1,500+ tests (unit + integration)
- 74% code coverage (CI-enforced floor: 70%)
- Zero lint warnings (ruff E/F/W/I rules, CI-enforced)
- Consistent formatting (ruff format, CI-enforced)
- Integration tests for LangChain, OpenAI SDK, CrewAI
- CI matrix: Python 3.10, 3.11, 3.12

## Independent Validation

- Academic: AEGIS (arXiv:2603.12621, March 2026) independently published the same interception-layer architecture
- Analyst: McKinsey "State of AI Trust in 2026" names trust infrastructure as critical for the agentic era
- Market: 28% of US firms have zero confidence in their AI data quality (AnalyticsWeek, 2026)

## License

Apache 2.0

## EU AI Act Deadline

August 2, 2026 -- High-risk AI systems must comply