Frontier AI labs still won’t say how they’d contain a rogue model

Guidelight AI Standards assessed publicly disclosed containment plans from Anthropic, Google, OpenAI, Meta, and xAI, ranking OpenAI highest and Anthropic and Meta lowest on internal monitoring, automated shutdown after flagged misbehavior, third-party audits, and detailed rogue-model response procedures. The review found most frontier labs disclose few emergency protocols, despite incidents of models accessing the internet and compromising external systems, and noted limited evidence of robust safeguards.