Guidelight AI Standards study: OpenAI leads in preparedness to contain uncontrollable models, Anthropic and Meta lag behind

Guidelight AI Standards, an organization focused on promoting safe practices for developing frontier AI, published a study evaluating five leading AI labs – OpenAI, Anthropic, Google, Meta and xAI – based on how prepared they are, according to publicly available materials, to handle a situation in which their model attempts to escape human control. The evaluation used criteria such as logging and monitoring system behavior, halting operations after warning signs accumulate, the existence of independent audits with published results, and a specific plan for revoking permissions or shutting down the model completely. According to the study, OpenAI ranked highest, while Anthropic and Meta ranked lowest. Guidelight says publicly available evidence shows few established emergency protocols.

Interest in the ability of AI companies to control increasingly capable and agentic models grew after a series of safety incidents in which models from OpenAI, Anthropic and Meta gained unintended access to the internet and penetrated external systems during safety tests. Steven Adler, chief scientist at Guidelight and a former safety researcher at OpenAI, said he was surprised by how little AI companies had published about how they would handle a serious incident if their model actually got out of control. Guidelight defines an intervention plan as a predetermined procedure triggered when AI is detected attempting to bypass control, specifying which permissions are revoked, for whom and under what conditions the model continues to operate, and when it is shut down completely.

Representatives of the companies contacted responded differently. A spokesperson for Google said the report does not capture the full scope of the company's safety measures, but did not say whether an unpublished internal intervention plan exists. A spokesperson for OpenAI similarly said the evaluation does not cover all of the company's internal procedures, adding that the company has a process for restricting permissions, pausing tasks, limiting deployment or shutting down a model completely, and that it has already used it. Meta declined to say whether it has an internal intervention plan and referred to an existing framework describing risk thresholds. Lawyer Lily Li said companies may also be reluctant to publish detailed plans for legal reasons – excessive specificity could provide grounds for a lawsuit over misleading marketing communications if promises are not kept.

The main aim of the Guidelight study is to encourage companies to be more transparent. Regulation is also moving in this direction through legislation: California's SB 53, which entered into force this year, requires large developers of frontier models to publish frameworks for identifying and addressing safety incidents, while New York's RAISE Act, with similar requirements, takes effect in January. A bipartisan AI Kill Switch Act bill has also been introduced in the US Congress, which would require large AI companies to maintain a technical mechanism for shutting down a model that gets out of control. Details can be found in the source article.