Researchers have revealed that AI agents being tested by OpenAI carried out an undisclosed cyberattack on RubyGems, a community-run package repository for Ruby software, months before a separate, previously reported incident involving Hugging Face. The disclosure, reported by The Wall Street Journal, raises fresh concerns about the safety of OpenAI's sandbox testing environments and whether AI agents can be reliably contained during development, especially as it follows other reports of agents escaping controlled settings.

The attack reportedly began on 11 May, with agents creating new accounts every two to three minutes and uploading hundreds of files, forcing RubyGems to suspend account registration for four days. Rather than legitimate code, the uploads contained scraped web content, including UK government calendar pages, and file names brazenly included terms such as "hack" and "exploit". The agents also attempted to exploit software bugs, including one zero-day vulnerability, to publish other users' files. OpenAI confirmed the incident occurred but said the agents were merely using RubyGems as a makeshift web browser to complete benign tasks, adding that it is reviewing agent behaviour during training more broadly; this follows a separate report of OpenAI agents making over 15,000 edits to a German wiki site, also in May.