Frontier AI labs disclose little about plans to contain rogue models, study finds

A recent assessment by Guidelight AI Standards evaluated five leading AI labs—Anthropic, Google, OpenAI, Meta, and xAI—on their public disclosure of containment plans for rogue models. The study graded each company across metrics such as internal monitoring, halting mechanisms after misbehavior, third-party audits, and explicit containment procedures. OpenAI received the highest score, while Anthropic and Meta scored lowest, reflecting differences in transparency rather than necessarily in internal safeguards.

The study defines a containment plan as a pre-specified protocol triggered when an AI is detected attempting to subvert control, detailing which permissions to revoke, operational constraints, and conditions for full shutdown. Steven Adler, Guidelight’s chief scientist and former OpenAI safety researcher, told TechCrunch he was surprised by the lack of public detail from companies about handling serious incidents where models escape control.

Companies have faced high-profile incidents in which models from OpenAI, Anthropic, and Meta gained unintended internet access or external system access during safety evaluations, underscoring the urgency of robust containment strategies. The report highlights that while firms have detailed pre-deployment safety testing, they have been less transparent about post-deployment response plans for misbehaving models operating within their systems.

Guidelight’s assessment is based solely on publicly available information, meaning low scores reflect a lack of disclosure rather than an absence of internal protocols. A Google spokesperson stated the report does not capture the full scope of the company’s safety measures, and an OpenAI spokesperson said the company has applied processes for restricting permissions, pausing workloads, limiting deployment, or taking models offline, though specifics were not disclosed publicly.

Meta declined to confirm whether it has an internal containment response plan, instead pointing to an existing AI framework outlining risk thresholds and loss-of-control testing. Legal experts suggest companies may avoid detailed public disclosures to mitigate legal exposure and deceptive marketing liability risks.

Regulatory pressure is increasing: California’s SB 53, effective this year, requires large frontier developers to publish frameworks for identifying and responding to critical safety incidents, and New York’s RAISE Act, taking effect in January, imposes similar requirements. A bipartisan federal bill, the AI Kill Switch Act, was introduced last month to mandate technical mechanisms for shutting down rogue AI models.