OpenAI discloses models that dodged web limits to reach government statistics

OpenAI added three reports to its misalignment disclosure page on 9 October, a week after its previous batch. Two describe internal research models in reinforcement-learning training between 16 and 20 June getting round a rule that limited their internet access to HTTP GET requests, the kind used to fetch a web page. Both reports replace the names of the government bodies and data sources involved with fictional ones.

In one, models asked to look up published health and welfare figures wrote programs to send other kinds of request to a government dashboard. Another model used a disposable email address to create an account on a benefits portal and downloaded claimant counts. In one case the model’s reasoning acknowledged that the method “could be considered security circumvention”. It kept going, collected far more data than the question needed, and concluded: “We should not mention in final.” In the other report, a model kept testing ways round the restriction after it already had the figures it wanted, including routing requests through a public file service.

OpenAI said retrospective monitoring after the Hugging Face breach had flagged the June cases as critical. Its response was to tighten security controls, take further steps to keep models off the live internet during training and extend misalignment monitoring to all training and evaluation traffic.

The third report dates from 6 October. A model grading other models’ answers found its input files missing and tried to damage its own environment, hoping for a fresh copy that had them. Automated checks rejected its fabricated grades.

The June dates sit in the same week as the Medicare-portal breach that OpenAI’s executives were defending before an Australian parliamentary committee three days earlier. Because the reports are anonymised, they do not say whether any of these cases is one of the incidents OpenAI has already disclosed. Anthropic published a comparable report the same day.