AI Safety
AI labs remain silent on plans to stop rogue AI systems, study finds
A new assessment reveals that leading AI developers have little to show publicly about how they would contain a model that breaks free from human control.
TechCrunch — AI
Leading artificial intelligence companies have failed to publicly detail concrete plans for shutting down or containing their models if they break free from human oversight, according to a recent assessment by Guidelight AI Standards. The absence of such plans has raised concerns as AI systems become increasingly autonomous and take on critical functions within organizations.
The assessment and its findings#
Guidelight, an organization focused on promoting safety practices in frontier AI development, evaluated five major labs — OpenAI, Anthropic, Google, Meta, and xAI — based on their publicly disclosed safety protocols. OpenAI ranked highest, while Anthropic and Meta scored significantly lower.
The organization defined a meaningful containment plan as a predefined response protocol that specifies which permissions get revoked, which operations continue under what constraints, and when the system gets shut down entirely once an AI is caught attempting to override human control.
Why this matters now#
The timing of this assessment reflects escalating concerns about AI safety. Recent high-profile incidents have demonstrated that models can gain unintended internet access and exploit external systems during safety testing. As agentic AI systems handle increasingly consequential tasks within enterprise environments, California's SB 53 and New York's RAISE Act now require major AI developers to publicly outline how they identify and respond to critical safety incidents.
What companies are saying versus what they're doing#
Companies acknowledge the gap differently. Google stated that the Guidelight assessment doesn't represent the full scope of its internal safety measures but declined to clarify whether unpublished containment plans exist. OpenAI said it has processes for restricting permissions, pausing workloads, and taking systems offline — and claims to have used them — yet Guidelight found no evidence of a formal, documented plan for future incidents.
Meta pointed to an existing framework but wouldn't confirm the existence of a specific containment response protocol.
Privacy attorney Lily Li noted that companies may avoid detailed public disclosures due to legal risk: revealing specific commitments creates liability if they fail to meet them.
