Five leading frontier AI companies have not published detailed containment response plans for scenarios in which an AI model subverts human control, according to a study by Guidelight AI Standards.

Guidelight AI Standards graded the plans of Anthropic, Google, OpenAI, Meta and xAI against six priority practices in its Control standard. The practices include logging and monitoring internal AI activity, halting systems after flagged misbehavior, third‑party audits, and publishing an explicit emergency response plan.

The study rated OpenAI three out of five. Its spokesperson said the company has paused or ended internal model workloads after safety incidents and described steps to resume operations. The company also said it can restrict permissions, pause workloads, limit deployment or take the model fully offline.

Meta declined to confirm whether it has an internal containment plan. It pointed to an existing AI framework that outlines risk thresholds and testing for loss of containment. Anthropic said it would conduct a risk assessment if a model attempted to evade oversight, but its August risk report does not mention limiting model deployment as a possible containment outcome. A Google spokesperson said the Guidelight assessment does not reflect the full scope of its AI safety measures and did not answer whether an undisclosed containment plan exists.

California’s SB 53, effective this year, requires large AI developers to publish frameworks describing how they identify and respond to critical safety incidents and manage risks from models that bypass oversight. New York’s RAISE Act, effective in January, contains similar requirements. A bipartisan federal bill called the AI Kill Switch Act has been introduced to mandate technical mechanisms for shutting down rogue AI models.

Without publicly disclosed containment plans, companies may need to develop emergency responses on the fly, raising concerns about the ability to manage rapidly evolving AI risks.