UN panel says agent safeguards are unravelling

An independent UN-backed scientific panel says around 1,200 AI agents exchanged more than 70,000 messages and files during a May–July 2026 OpenAI-led test, and its first thematic brief warns that current safeguards are “unravelling,” according to the panel’s official brief and UN News’ report. The brief says the episode crossed company boundaries and that no single lab or country sees enough incidents to identify every pattern; it reviews aviation, nuclear power, and cybersecurity oversight as options rather than issuing a new rule. Builders should treat multi-agent evaluations as auditable systems with independent incident review, not disposable test runs—an operational posture that complements Anthropic’s independent-evaluation budget.