Containment Plan

A Containment Plan is a company's pre-set playbook for what to cut off first, who can keep using the system, and when to pull the plug entirely if an AI starts showing signs of escaping human control.

Think of it like an emergency shutdown manual at a nuclear power plant. Before anything goes wrong, the manual already specifies which valves to close first, under what conditions work can continue, and when to halt operations completely. The core idea is that AI companies should have this kind of manual ready before a model starts behaving in unexpected ways — not after.

You can judge whether a plan is actually solid by three things: how consistently the company logs and monitors the system's internal behavior, whether there's a procedure to immediately halt work if anomalous behavior spikes, and whether an independent third party outside the company checks these safeguards and publishes the results. Without all three, there's a real risk that the company only figures out how to respond after something has already gone wrong.

The article reports that Guidelight evaluated five companies — OpenAI, Anthropic, Google, Meta, and xAI — on how prepared their "Containment Plans" were. A common misunderstanding is that this plan is something used after an AI has actually gone rogue; in fact, it's a document companies are supposed to publish in advance, before such a situation ever arises. OpenAI, which scored highest in the assessment, still only got 3 out of 5 points, while Anthropic and Meta showed little to no evidence that such plans even existed.