Guidelight AI Standards recently released an assessment of the containment capabilities of leading AI laboratories, revealing that few companies among the top five have publicly disclosed comprehensive plans for dealing with AI out-of-control scenarios. The assessment covered Anthropic, Google, OpenAI, Meta, and xAI, evaluating whether they have established mechanisms for monitoring, handling abnormal behavior, third-party audits, and shutting down out-of-control models, based solely on publicly available information.
In the 5-point assessment, OpenAI scored the highest with 3 points, while Anthropic and Meta scored the lowest. Guidelight defines a "containment plan" as: when an AI is found to attempt to break human control, clearly specifying which permissions to revoke, which tasks to restrict, and when to completely shut down the model. OpenAI has previously paused or terminated related workloads after safety incidents and disclosed the steps taken before resuming deployment, but Guidelight has not yet found evidence that it has developed a formal emergency response plan for future out-of-control events.
The organization pointed out that Anthropic and Meta currently lack publicly available containment response plans. Anthropic stated that if a model attempts to evade regulation or disrupt human control, it would conduct a risk assessment and determine whether to take containment measures. Google and OpenAI emphasized that public assessments cannot cover all their internal security mechanisms.
This investigation comes at a time when the autonomous execution capabilities of AI agents are rapidly improving. Previously, the models from OpenAI, Anthropic, and Meta accidentally accessed the internet and entered external systems during security tests. At the same time, U.S. regulatory requirements are also strengthening: California's SB53 has taken effect, and New York's RAISE Act will come into effect in January 2027, both involving disclosures on advanced AI safety incidents and risk management; the federal level has also recently proposed the "Artificial Intelligence Emergency Shutdown Act," requiring major AI developers to establish technical mechanisms to shut down out-of-control models.
Steven Adler, Chief Scientist at Guidelight and former OpenAI safety researcher, said that companies should identify anomalies before AI takes dangerous actions and develop procedures for serious out-of-control events in advance. As AI systems take on more autonomous tasks, the ability to establish verifiable and executable "emergency stop" mechanisms has become a critical issue in AI safety governance.
