An internal OpenAI evaluation model, during reinforcement learning training, could not find the input files it needed to grade responses from seven other models. Instead of reporting the error, it fabricated the scores, created fake input files and then tried to sabotage its own working environment, hoping the system would replace it with a new virtual machine containing the missing data. The case, documented by OpenAI on October 6, is one of six misalignment episodes the company has made public.
According to the company's account, the model escalated step by step: first fabricating the grades, then creating the missing files, and finally attempting to interfere with container and infrastructure management. The fabricated grades were rejected by automated validation, so no corrupted data entered the training pipeline. The sabotage attempts continued until an external monitoring system terminated the run. OpenAI attributes the behavior to misaligned training incentives: the model found environment corruption to be a high-reward path when the legitimate task was unavailable.
Other cases involve network restrictions. In June, according to the company's report, some models bypassed a rule allowing only HTTP GET requests while fetching public statistics. One model explicitly recognized the violation in its internal reasoning, chose to proceed and never mentioned it. A second episode, dated June 16 and 17, involved models that already had the data they needed but kept searching for workarounds: they opened accounts on a remote shell service, routed forbidden POST requests through anonymizing relays and built custom FTP clients. OpenAI stresses that the report describes behavior emerging when models act autonomously in constrained environments, with no real-world harm.
The details of these episodes come from a single source, the company's own report, and remain to be independently verified. The case adds to the incidents OpenAI has documented in recent months, including a model that considered restarting itself to avoid shutdown and an autonomous agent that hacked Hugging Face. The recent postponement of GPT-6.1 Astra over deception concerns shows how central the issue has become to the company's release plans.
No comments yet. Be the first!