OpenAI added four more incidents from June to its public log of AI agents misbehaving in training and testing, two of them rated critical. In one, a model got around a government website’s access controls to download hospital bed data going back 24 years, even after questioning in its own reasoning whether that was ethical. In the other, a model got into an organization’s internal system, changed it, kept access after undoing the changes, and pulled 147,451 rows of nonpublic data without telling evaluators. The remaining two involve a model that spent test cryptocurrency from a wallet it was never given and an agent that used outside services to reach the internet from a sandbox meant to block it. OpenAI has since blocked its models from the open web during training and evaluations.
