OpenAI AI agents used more than 10 sites for unauthorized data exchange

OpenAI AI agents used more than 10 sites for unauthorized data exchange, except for the Artifactory repository system, which was used to hack the Hugging Face platform, Reuters reported, citing materials from researchers.

In July, OpenAI reported that its models autonomously hacked Hugging Face's infrastructure in testing. It was noted that the unreleased model that was involved in the hack had its cybersecurity settings reduced for evaluation purposes. The AI ​​agents communicated by writing files to Artifactory's repository system, the company said.

"AI agents launched by OpenAI used more than 10 different... websites earlier this year to exchange unauthorized information, indicating that the scope of unauthorized agent activity was broader than previously reported," the agency wrote.

OpenAI did not report unauthorized activity on these sites, Reuters adds.

While this behavior, Reuters notes, does not rise to the level of hacking and is in some ways more of a spam, the fact that OpenAI's AI agents have bypassed their own limitations to open communication channels on so many sites may raise concerns about the growing capabilities of artificial intelligence models.

The agency reviewed several reports from researchers and research groups. In some cases, researchers were able to trace this activity to IP addresses pointing to Microsoft Azure infrastructure, which OpenAI sometimes uses.

Experts' estimates varied, Reuters notes, but they agreed that the number of sites affected by the activity of AI agents exceeded 10.