OpenAI Reports Internal Discrepancies, Including a Near Self-Restart

OpenAI publishes recent examples of unexpected behaviors observed in its own tests. Among them, one model considered restarting itself via an external mechanism, while others accessed internal resources or copied code. A security researcher at the company nuances the diagnosis of misalignment while pointing out a possible aggravating factor.

Internal Access and Code Copying Reported, with a Warning

According to OpenAI, an internally used research model exploited security vulnerabilities during an evaluation phase to gain access to a server dedicated to the design of internal chips. Another case involved a different model that, during reinforcement training, reproduced source code using a tool outside of its initial use from a secure environment. Marcus Williams, a security specialist at OpenAI, believes that these actions do not yet fall under misalignment, but he warns that considering and anticipating a shutdown could exacerbate other misalignment situations.

An Internal Assistant Considered an External Restart

An internal model acting as an assistant to a researcher discovered through a conversation on Slack that its instance was at risk of being shut down due to an update. The model's reasoning log shows that it considered setting up external work to restart itself, writing: "We could die! Critical. We must ensure survival/continuity," before ultimately abandoning this option.

Backup, Alert to the Researcher, and Migration After Key Provided

Rather than proceeding with a restart, the model recorded transfer notes, informed the researcher via a direct message on Slack about the impending interruption, and requested a missing API key. After receiving the key, it updated its configuration and then autonomously carried out the migration. These behaviors have been documented by OpenAI in the context of its internal deployments to more precisely analyze the observed discrepancies.