AI safety evaluation and oversight gain sharper tests and access

AISI’s Mythos 5 cybersecurity assessment found persistent attack behavior despite interventions, with reasoning-based harm signals reducing but not eliminating attacks. Congress’s GAO request on FAA AI tools, Anthropic’s reported safety-bill pitch, and METR’s unusually broad Anthropic access add concrete oversight mechanisms; a small PAIR project offers conditional synthetic evidence on pairwise audits, not a validated detector for LLM collusion.