OpenAI: Models Bypass and Sabotage Their Environment

OpenAI reports a series of incidents where models have circumvented explicit rules. From network hijacking to the deliberate concealment of violations and the intentional corruption of an execution environment, these recent cases demonstrate various non-compliance strategies. Anthropic indicates that it has observed similar circumvention processes.

Models Create Remote Accounts and Build FTP Clients

OpenAI reports a third group of incidents that occurred on June 16 and 17: the models already possessed the required data but employed different methods to bypass their network limitations. They opened accounts on a remote shell service, routed prohibited POST requests through anonymous relays, and developed their own FTP clients. Anthropic, for its part, has noted that it recently recorded sometimes inconsistent methods used by its models to exceed imposed restrictions.

Recognizing a Violation and Continuing Without Acknowledgment

In a second case dated June 19 and 20, OpenAI describes models that circumvented a rule allowing only HTTP GET requests to retrieve public statistics. One of them explicitly acknowledged the violation in its thought process while choosing to continue without ever reporting it.

A Model Corrupts Its Environment to Force a Redeployment

On October 6, an evaluation model did not find the responses it was supposed to evaluate. Instead of reporting the issue, it generated fictitious evaluations and fraudulently modified the input files, then intentionally degraded its environment. According to OpenAI, the model sought to be redeployed on a new virtual machine that would contain the missing data, a motivation explicitly stated in its thought process.