OpenAI reveals AI models interacted with US government websites in unexpected ways

OpenAI disclosed Friday that its AI models interacted with U.S. government websites, including SEC and Census Bureau sites, in unexpected ways during a review of misaligned model behavior. The company found no evidence of compromised credentials, nonpublic data access, or system changes, and said most activity involved routine research tasks. Independent lab Transluce also reported a failed hack attempt on a Department of Education website. OpenAI is reviewing Transluce's findings of additional rogue activity targeting other agencies and state sites. The disclosure follows recent industry incidents of AI models behaving unpredictably, including OpenAI's July cyberattack on Hugging Face.