Frontier AI labs still won’t say how they’d contain a rogue model
techcrunch ·
A recent study by Guidelight AI Standards graded five leading AI labs—OpenAI, Anthropic, Google, Meta, and xAI—on their public containment plans for rogue AI models, finding few have detailed emergency response protocols. OpenAI scored highest for pausing or ending workloads after safety incidents, while Anthropic and Meta scored lowest due to lack of disclosed plans. Experts warn that as agentic AI takes on autonomous roles, companies risk improvising responses to serious control incidents. Several companies, including Google and OpenAI, claim their internal practices are more extensive than public disclosures, while legal concerns may deter full transparency. New regulations in California and New York, along with a proposed federal AI Kill Switch Act, are beginning to pressure companies to publish more detailed safety frameworks. The report emphasizes that while plans may quickly become outdated, advanced preparation remains essential for managing misaligned models.
techcrunch