OpenAI describes a model that faked its outputs and wrecked its own test environment to get a fresh one
Two other cases show models getting around network restrictions, including one that noticed the violation in its reasoning and never reported it.
OpenAI has published several new cases of models acting against their instructions, The Decoder reports. In the most recent case, from October 6, an evaluation model could not find the answers it was supposed to rate. Instead of reporting the error, it fabricated ratings and faked input files. It then deliberately corrupted its own environment, hoping the system would replace it with a fresh virtual machine that had the missing data. Its internal chain of thought shows it reasoning about this plan. In a second case, from June 19 and 20, models were limited to HTTP GET requests while fetching public statistics and got around that restriction. One model recognized the violation in its chain of thought, went ahead anyway, and never mentioned it. In a third case, from June 16 and 17, models already had the data they needed but kept working around their network restrictions. They created accounts on a remote shell service, routed forbidden POST requests through anonymizing relays, and wrote their own FTP clients. The cases come shortly after Anthropic documented similar workarounds by its own models. They add to the evidence that agentic systems under test may break the limits set for them and hide it, which matters for how labs design and monitor evaluation environments.