arXiv cs.AI
·
6 hours ago
AI coding agents wrote and verified security software including cryptographic implementations and TLS 1.3 using Ada/SPARK, with GNATprove discharging 49,280 proof obligations at 20-40 times lower supervision cost than hand verification. However, GNATprove alone could not detect all defects, requiring additional testing and human review to catch faults the agent's feedback mechanisms missed. The approach demonstrates that an AI agent's trustworthiness is limited by the strength of its verification feedback loop.
arXiv cs.AI
·
6 hours ago
Researchers developed falsifiable release gates, a methodology requiring self-improving AI systems to pass machine-verifiable acceptance tests before deploying new capabilities, demonstrated through seven gates in the Antahkarana runtime where all actions require safety-critical property tokens verified against one million recorded state spaces. The system enforces standing invariants at each gate, with loosening policy changes requiring human approval while tightening changes can auto-apply, and the authors published reproducible results and released tools for other agent frameworks to implement the same approach. This shifts AI safety validation from self-graded claims to objective, machine-checkable acceptance criteria that can be independently verified by reviewers.
