A CTO opened a pull request defining six autonomous AI agents (five of them instructed to modify code in his repository on their own) and merged it eleven seconds later. Plus 2,593 lines, zero human reviews, blank checklist. This is not an outlier.

AI agent frameworks ship oversight, orchestration, and verification tooling to production teams. Their engineering processes don't follow their own rules. This isn't hypocrisy. It's structural, and it's measurable.

I built a forensic static analysis engine for AI-augmented codebases: polyglot SAST inside an air-gapped container, cryptographic attestation via the AIV (AI Integrity Verification) protocol, and a per-blank human verification gate that produces a defensible audit corpus. I've used it to audit five AI agent frameworks from public repository data. Forty-four verified findings. Every one anchored to a specific commit a reviewer can verify in thirty seconds.

In this talk I'm running the engine live, on stage, on a popular AI agent framework.

You'll watch:

• A static analysis container start with no network access, write a cryptographically-chained destruction record, and produce structured findings in real time.
• The engine reject one of its own machine-generated findings because the evidence quote doesn't byte-match the source. AI defensibility, demonstrated.
• A per-claim human attestation gate where the operator clicks through structured verifications, building cognitive evidence one accepted blank at a time. This is the moat, and you'll see why it can't be automated away.
• A board-ready audit briefing rendered live, GPG-signed, ready for delivery.

Then the methodology takeaway: five questions every defender should be able to answer about any AI-augmented codebase before it touches production, drawn from the patterns I found across five framework audits. Plus three concrete numerical checks anyone can run on their own GitHub repos Monday morning.

The pattern across the audits: AI-augmented velocity creates verification debt faster than teams can pay it down. The attack surface is not the generated code. It is the process that produced it.
This is a live demonstration of forensic technical due diligence, a practice that's emerging because AI codegen has outpaced the governance meant to oversee it.