Notes from the evidence layer.

How NemoGuard NIM classifier verdicts become signed detector findings in the AIR evidence chain.

The state of AI agent accountability in 2026: the incidents, the converging regulatory deadlines, and the evidence infrastructure gap.

Evidence-grade infrastructure for accountable AI agents: signed intent capsules, 16 OWASP-mapped detectors, causal explanation, Auth0 containment, and cross-agent chain of custody.

Structural Verification: a deterministic floor that checks whether the agent’s actual trajectory served its declared intent, and cannot be prompt-injected.

The HIPAA Security Rule NPRM makes audit controls mandatory. AI agents touching PHI need cryptographic evidence chains, not application logs.

NVIDIA NemoClaw controls what clinical AI agents can do. Project AIR proves what they did. Prevention plus evidence for regulated healthcare.

A map of AI agent security tooling, and the post-incident forensic layer most teams don’t realize they’re missing.

Anthropic’s paper on trustworthy agents names three ecosystem gaps. Project AIR is our answer to evidence sharing and open standards.

From pip install projectair to a signed forensic timeline of your LangChain agent in under five minutes.

What post-market monitoring evidence actually has to contain by August 2, 2026, and how to automate producing it.

92% of MCP servers lack proper OAuth. We scanned real configurations and found critical vulnerabilities across authentication and authorization.