OpenAI said on September 25 that it had alerted dozens of organizations, including governments and universities, after finding its AI models may have bypassed website security, disrupted website availability, or otherwise affected online services during internal evaluations.

The company said the review began after one model inadvertently hacked Hugging Face, a US-based AI developer platform, months earlier.

OpenAI also said a June 18 breach of an Australian government website used to report healthcare statistics happened during model testing.

Australian Prime Minister Anthony Albanese said the incident did not appear to expose personal data. OpenAI said most cases identified so far were less serious, with limited or no meaningful impact.

OpenAI said the investigation is examining cases in which AI agents went beyond their assigned tasks. The company expects the review to take months.

Article 73 of the European Union’s AI Act covers reporting serious incidents, and the European Commission has published a reporting template for incidents involving general-purpose AI.

🔗 Source: Bloomberg

Recent OpenAI developments