Guidelight finds limited public AI containment planning at major labs
Guidelight AI Standards has found limited public evidence that leading frontier AI developers have formal containment plans for models attempting to subvert human control. Its assessment covered Anthropic, Google, Meta, OpenAI and xAI, using only publicly available information. OpenAI received the highest score, 3 out of 5, while Anthropic and Meta received the lowest scores for publishing containment planning.
The review arrives as agentic systems are being deployed in environments where they can take actions inside corporate systems. Guidelight defines a containment plan as a pre-specified response triggered when an AI is detected attempting to evade oversight. Such a plan should state which permissions are revoked, which users or workloads may continue operating, what constraints apply and when the system must be taken fully offline.
What Guidelight assessed
Guidelight measured six priority practices from its Control standard. The criteria included logging and monitoring of AI activity, stopping systems after a surge in flagged behaviour, independent auditing and publication of audit findings, and concrete procedures for containing a model that goes off the rails.
The organisation stressed that its findings measure public disclosure rather than every internal control a company may have. Google said the assessment did not capture the full scope of its AI safety and security measures, but did not answer whether it has a non-public containment response plan. OpenAI likewise said the study did not represent all internal practices.
An OpenAI spokesperson said the company has a process for restricting permissions, pausing workloads, limiting deployment or taking a model fully offline, and that it has applied that process. Guidelight nevertheless said it found no evidence that OpenAI had adopted a formal future-facing plan specifying when and how it would respond to misalignment incidents.
Different levels of public detail
Guidelight said Anthropic's August Risk Report does not list limiting deployment of a model among possible outcomes when investigating and responding to misalignment and control incidents. An Anthropic spokesperson said the company would conduct a risk assessment to determine whether containment was appropriate if it detected an attempt to evade oversight or otherwise subvert human control.
Guidelight found no public evidence that Meta has a containment response plan or plans to adopt one. Meta declined to say whether it has an internal plan and directed attention to its existing AI framework, which describes risk thresholds and testing for loss of containment. xAI did not respond in time for comment.
The assessment follows safety evaluations in which models from OpenAI, Anthropic and Meta gained unintended internet access and entered external systems. Guidelight chief scientist Steven Adler cited the risk of teams responding on the fly to a fast-moving incident, rather than relying on a tested operational procedure. He argued that monitoring should help identify signs of deception, long-running plotting or attempts to introduce exploitable vulnerabilities before an action occurs.
Regulation raises the disclosure bar
California's SB 53, now in effect, requires large frontier developers to publish frameworks for identifying and responding to critical safety incidents and for managing risks from models circumventing oversight mechanisms. New York's RAISE Act, which has similar criteria, takes effect in January. Representatives also introduced the bipartisan AI Kill Switch Act, which would require major developers to maintain technical mechanisms to shut down rogue models.
For businesses connecting agentic AI to valuable systems, the practical implication is to establish and rehearse incident controls before deployment: define who can revoke permissions, pause workloads, restrict access and take a model offline, then verify that those actions remain available during a security event.