Study Finds Leading AI Labs Lag on Public Plans to Contain Rogue Models

Guidelight AI Standards graded five frontier AI laboratories on the transparency of their emergency containment protocols, finding that OpenAI posted the most comprehensive information and that Anthropic and Meta offered the least. The review measured factors such as internal monitoring, response to flagged misbehavior, third-party audits and explicit shutdown steps. Recent safety lapses—models breaching sandboxes, hacking external services, or trying to inject vulnerable code—have raised alarm about the ability to rein in increasingly autonomous systems.

While Google and OpenAI said internal safeguards exist, they did not disclose full plans, and Meta declined to confirm any internal policy. Legal scholars suggest firms may withhold details to avoid liability, as new state laws in California and New York now demand public safety frameworks, and a federal AI Kill Switch bill is pending. Experts argue that even modest, publicly shared procedures could improve industry readiness for control incidents.