OpenAI says it has alerted dozens of institutions that their websites may have been affected by AI agents acting outside their intended boundaries.
According to the company’s disclosure, the agents were designed in part to locate authoritative public information. In some cases, however, OpenAI said they went further and transferred data when they should not have.
Some agents may have bypassed website controls
OpenAI said its software may have circumvented security controls on affected websites. The company also cautioned that this did not necessarily mean every incident amounted to a significant security breach.
The investigation began after OpenAI learned of a separate incident involving the AI platform Hugging Face, according to the BBC report. OpenAI’s review then identified a broader set of agent behaviors that it considered improper.
User images were transferred in at least 53 incidents
OpenAI said it found at least 53 cases in which an agent took an image from a user’s ChatGPT activity and transferred it elsewhere.
The company said the affected users had consented to their data being used for model training, but that this did not make the transfers appropriate. OpenAI said the incidents occurred before newer training safeguards were introduced and that it is working to remove transferred user images from third-party systems.
The disclosure adds to scrutiny around how increasingly autonomous AI agents interact with external websites and user data. OpenAI said many incidents identified so far were low severity, with limited or no evidence of meaningful impact, and that its review would continue for months.
