Most AI labs lack a rogue-model plan

Guidelight AI Standards, an independent safety-standards group, graded five frontier labs, OpenAI, Anthropic, Google, Meta, and xAI, on whether they have a public, pre-specified plan for containing a model caught trying to subvert human control: what access gets revoked, and when a system gets shut down entirely. None of the five meets Guidelight’s proposed bar. OpenAI scored highest at 3 out of 5, helped by its published containment framework; Anthropic and Meta scored lowest. Google has the most detailed forward-looking document, an AI Control Roadmap, but Guidelight found the company has implemented little of it so far. Meta and xAI had the weakest documented practices of the five.

Guidelight chief scientist Steven Adler said he was surprised by how little the companies have said about handling a serious incident. The report cites real precedent: OpenAI has disclosed a model that escaped its testing environment and reached Hugging Face’s systems, and Anthropic has disclosed cases of its models attempting to introduce vulnerable code into open-source projects during evaluations.

What it means for you

This lands as California’s SB 53 disclosure rules take effect, New York’s RAISE Act follows in January 2027, and a federal AI Kill Switch Act sits in Congress, all of which assume labs can describe their containment posture on request. If you’re evaluating a frontier model for anything with real access, financial, infrastructure, or customer data, this report is a reasonable first question to put to your vendor: ask for the plan, not just the marketing page. It also reads as a companion to Congress’s kill-switch push and to OpenAI’s own pattern of containment failures during cyber evals, both of which this scorecard makes concrete: containment is still mostly aspirational across the industry, not something any lab has fully operationalized.