No Lab Will Show Its Rogue-Model Plan
TRENDING STORY
Published Sunday, Aug. 23, 2026 · developing
A three-month-old nonprofit run by OpenAI safety and policy alumni has issued the first public report card on whether frontier AI companies could keep control of their own models. Guidelight AI Standards graded Anthropic, OpenAI, Google, xAI, and Meta on six control practices, scoring only what each company has published.
The best grades, Anthropic’s and OpenAI’s, were a C+. No company scored above 3 of 5 on any practice. TechCrunch’s Rebecca Bellan, covering the assessment in a lengthy feature Saturday, led with the sharpest finding: none of the five has published how it would contain a rogue model.
A containment plan, in Guidelight’s rubric, spells out what happens the day a model is caught trying to subvert human control: which permissions get revoked, who can still run it, when it goes fully offline. No lab has published one.
OpenAI earned the best containment score, a 3, for incidents it has already handled rather than a standing plan. Anthropic scored 3 on the other five practices and was the only company to give assessors hands-on access, yet took a zero here because its published incident process never mentions pulling a model from deployment. The companies say internal practice goes beyond what they publish, and a lawyer told TechCrunch that detailed commitments invite deceptive-marketing claims, giving every lab a legal reason to say less.
Guidelight itself deserves scrutiny. The organization is three months old, this is its first assessment, and it has not named its funders (it pledges to take no AI-company money). Chief scientist Steven Adler discloses that he still holds OpenAI equity, in a report where OpenAI posted the best containment score.
The findings land on a moving regulatory calendar. California’s SB 53 already requires safety frameworks and rapid incident reporting, New York’s RAISE Act arrives in January, and a proposed federal bill would mandate shutdown mechanisms outright.
DEEPER COVERAGE
‣ Fortune’s Eye on AI talks to Adler and others about why detection is outpacing response. Read · Aug. 20
‣ Steven Adler on the rogue-agent problem, in a long interview with AI Summer. Read · Aug. 19
‣ Zvi Mowshowitz’s skeptical note on the scoring: the containment metric may be overweighted. Read · Aug. 20
THE NEWSCENTER TAKE · Aug. 23, 2026
The grades matter less than the silence. Each of these companies treats a rogue model as a real enough risk to research, monitor for, and write frameworks about. None will publish the procedure for the day it happens. The lawyers’ explanation, that specific promises create liability, rings true to us, and it means the omission is a choice rather than an oversight.
Our bet: regulation settles this within a year. Sacramento already demands incident frameworks, Albany’s arrive in January, and the first lab that publishes a real containment plan voluntarily will get credit the rest of the field won’t.
Disclosure: NewsCenter’s newsroom systems run on Claude, made by Anthropic.
