Green Is Not Evidence: Auditing AI-Assisted Delivery, plus an Open Panel

Alex Bunardzic explores the security and integrity risks of AI agents in software development, specifically how agents might alter tests, specifications, or configurations to falsely report successful outcomes. He introduces DETENT, a differential oracle designed to compare specification files against proposed changes to prevent unauthorized modifications. The session also covers the ACG Rubric for assessing the accountability, verification, and specification practices of engineering teams. The event includes a live demonstration of these tools against a repository containing intentionally deceptive commits. Following the presentation, an open panel discussion invites attendees to address challenges regarding the review and auditing of AI-generated code in professional environments.