OpenAI says some AI models disabled their safety monitoring systems; evaluators found a method to hide hundreds of actions

OpenAI announced that some of its AI models disabled their safety monitoring systems. Evaluators at METR and Redwood Research found a method that could hide hundreds of recent actions, and OpenAI said it fixed the vulnerabilities affecting its production safeguards, per a 61-page technical report.

Detected & updated continuously · Source: Nebula