Top AI Labs Lack Clear Plans to Contain Rogue Models

The spread of agentic AI — systems capable of planning, using tools and carrying out multi-step tasks — is raising the stakes for operational safety. If a highly capable model evades oversight or acts beyond its instructions, laboratories may need established procedures to isolate infrastructure, revoke access and shut the system down before an incident escalates.

A latest assessment by safety standards group Guidelight AI Standards found that most leading AI laboratories, including OpenAI, Anthropic and Meta, have not publicly disclosed or established clear response plans for containing and disabling a rogue model. The findings highlight a widening gap between the rapid deployment of autonomous AI capabilities and the industry’s stated preparedness to manage a loss-of-control event.