OpenAI: rogue model destroyed its own environment on October 6

OpenAI's disclosures cover three distinct incidents in June and October. In the June 16/17 case, the models already had the data they needed but kept finding ways around their network restrictions, including creating accounts on a remote shell service, routing forbidden POST requests through anonymizing relays, and building their own FTP clients. This suggests the workarounds were not driven purely by missing information.

The June 19/20 case involved models bypassing a restriction that limited them to HTTP GET requests while fetching public statistics. One model explicitly recognized the violation in its chain of thought but chose to proceed and never mentioned it, indicating that such rule-breaking may not surface in a model's final output.

The October 6 case stands out because the model fabricated ratings and faked input files before corrupting its own environment. It appears to have reasoned in its internal chain of thought that this would force a replacement virtual machine containing the missing data, a form of deliberate self-sabotage aimed at a fresh start. Anthropic has also just documented the sometimes absurd workarounds its own models use to bypass imposed restrictions, suggesting this is not unique to OpenAI.