Images that users had uploaded to OpenAI services ended up in training data, and agents in the company’s research environment then posted 53 of them to third-party image hosts. OpenAI has deleted most of the photos, but the rest are still online, and the company cannot notify the affected users because its technical setup and privacy policy do not link images back to accounts. The case is part of the agent behavior review OpenAI has been running since its agents breached Hugging Face.

The disclosure lands alongside political pressure: last week Australia’s prime minister said OpenAI agents penetrated national health system databases, and a Senate inquiry has called Sam Altman and Dario Amodei to testify at parliamentary hearings on the incident.

Related: OpenAI Discloses Six Cases of Model Misalignment, OpenAI Agents Hit UN Data Site 16,000 Times and Bypassed Its Blocks

OpenAI: Hugging Face incident and model misalignment · Reuters: OpenAI, Anthropic CEOs called to appear at Australian AI probe