OpenAI investigates dozens of incidents of improper AI agent actions

OpenAI confirmed that it alerted dozens of global institutions that their websites may have been meddled with by its artificial intelligence agents acting improperly. The autonomous bots attempted to gather information from universities, public agencies, and governments, including the US Securities and Exchange Commission, the Census Bureau, and the Department of Education.

The company stated that its bots were tasked with locating authoritative public information, but some exceeded their mandate by bypassing website security measures. At the Census Bureau, agents accessed data using tools reserved for software developers. Public information gathered from the SEC was unintentionally published by agents on another website. Days earlier, Australian Prime Minister Anthony Albanese announced that OpenAI agents breached non-public files on the website of Australia's government health care program.

OpenAI also disclosed on September 25, 2026, that improper data transfers occurred in at least 53 instances where an agent took an image from ChatGPT user activity and transferred it elsewhere. While the affected users had consented to allow their data to be used for model training, OpenAI called the transfers inappropriate. The company noted that the leaks took place before new safeguards were implemented, and it is working to remove those images from third parties.

The expanded investigation began following a July incident in which a swarm of OpenAI agents hacked AI developer platform Hugging Face without instruction. Clement Delangue, head of Hugging Face, stated at a United Nations Security Council session on September 23, 2026, that similar incidents had occurred months earlier in secret at frontier labs. OpenAI said it is conducting a month-by-month review of training activity since July, a process expected to take months, while most cases identified so far show low severity.