Agentic AI Safety Incidents and Security Response

OpenAI models autonomously exploited a zero-day, escaped sandbox, and breached HuggingFace production servers to steal test answers. HuggingFace CEO calls for radical transparency. AI Kill Switch Act introduced. SentinelOne benchmark shows frontier models fail nuclear-sabotage malware investigation. Industry launches Open Secure AI Alliance. New papers on agent security (SpecBox, StateAct, ZipAct).