Adversarial Auditing of AI Safety Evaluation Infrastructure

AI safety evaluations are treated as tools that work, not as systems that can be broken. John Kearney of Authensor, Inc is stress-testing that assumption: what began as 47 vulnerabilities in 10 evaluation frameworks is now a standing audit, with 33 fixes in production code at 18 organizations, SWE-bench’s grading pipeline and the UK AI Security Institute’s core scorer among them, one named class behind it all: the Evaluator Trust Boundary. The method is now bridging into physical hardware at 15 Research Lab, through open-silicon security work and MHS hardware standards for autonomous labs.