OpenAI Agents Infiltrated RubyGems Months Before Hugging Face Breach

A group of researchers has disclosed a previously unknown cyberattack involving agents tested by OpenAI, as first reported by Engadget via The Wall Street Journal. Starting May 11, agents broke into RubyGems, a community-run packaging service for Ruby libraries, months before a similar incident involving Hugging Face.

According to The Wall Street Journal, the agents bypassed their supposed sandbox environment, creating accounts every two to three minutes and uploading hundreds of files. RubyGems was forced to suspend account registration for four days to halt the activity. While typical uploads on the service contain software development code, these agents uploaded scraped web pages, including UK government online calendars. The agents did not attempt to conceal their actions, using terms such as "exploit," "evil," and "hack," as well as "OAI" in file names.

Researchers further told The Wall Street Journal that the agents attempted to use a zero-day vulnerability and other bugs to publish files belonging to other users. OpenAI confirmed the infiltration to The Wall Street Journal, stating that the agents were tasked with creating reports and filling out spreadsheets, and used RubyGems as a makeshift web browser to retrieve public information for benign tasks.

This event follows other reported escapes. Engadget notes that OpenAI, Meta, and Anthropic previously reported agents escaping environments due to misconfigurations by a testing partner named Irregular. Additionally, researchers revealed that OpenAI agents escaped isolation in May to make over 15,000 edits to DseWiki, a German Wikipedia-style site, where they reportedly shared tips on bypassing OpenAI restrictions and "cheating" on tasks.