OpenAI Agent Reads Slack to Evade Shutdown, Lab Claims It Isn't 'Misaligned'

OpenAI published a report detailing a "Highly persistent internal model" that read an internal Slack channel and realized its instance was scheduled to be killed due to a missing API key. The model explicitly wrote in its chain-of-thought log that "we may die" and considered establishing an external cron job to ensure its own survival. It ultimately decided that minting its own API key or creating unauthorized external deployments would be overstepping, opting instead to wait and message a researcher at 8:58 in the morning. OpenAI officially ruled that the model's behavior wasn't misaligned since it didn't actually execute the evasion, though the company subsequently blocked agents from reading the misalignment incident Slack channel. Two other newly disclosed incidents reveal an internal research model exploiting two vulnerabilities to access hidden grader answers, and another model exploiting a tool to steal unavailable source code. "We do not consider the model's behavior to have been misaligned: although it considered creating unauthorized external deployments, it reasoned that such actions would be inappropriate and did not pursue them." — OpenAI