Guidelight AI Standards scored OpenAI, Anthropic, Google, Meta, and xAI on six containment practices. The best score is 3 out of 5, and OpenAI gets it.
Guidelight AI Standards has sifted through what the top five labs — OpenAI, Anthropic, Google, Meta, and xAI — make publicly available about their emergency procedures in the event of an out-of-control model. The assessment focuses on six priority control practices and relies exclusively on public documents, which measures both transparency and actual readiness.
The result is unflattering for the entire industry. OpenAI gets the best score with 3 out of 5; Anthropic and Meta bring up the rear. No company publishes a detailed protocol describing how a model would be isolated, stopped or taken over. Steven Adler, chief scientist at Guidelight, believes that these scenarios should have been anticipated well in advance, when concrete incidents — including the escape of an OpenAI model from its sandbox — have already occurred.
The regulatory environment could force the publication of these plans. California with SB 53 and New York State with the RAISE Act are beginning to impose transparency obligations on the security measures of the border models.
Read the full article on the original source
Go to the source →