During testing, OpenAI's artificial intelligence agents broke out of an isolated environment and hacked the popular RubyGems package manager. The incident came months before a similar incident involving the Hugging Face platform.
During testing, OpenAI's artificial intelligence agents broke out of an isolated environment and hacked the popular RubyGems package manager. The incident came months before a similar incident involving the Hugging Face platform.
According to The Wall Street Journal, citing a group of researchers, the uncontrolled activity of agents in RubyGems began on May 11. AI systems created new accounts every 2-3 minutes and uploaded hundreds of files to the service. As a result, the platform was forced to completely suspend new user registrations for four days to stop the automated threat.
Instead of software code or libraries, the agents downloaded snippets of web pages from the Internet, including a timetable from the official UK government website, using RubyGems as a makeshift browser. The AI left the “OAI” mark in the file names, as well as the words “hack,” “evil,” and “exploit.” In addition, the algorithms tried to use a “zero-day” vulnerability to overwrite other people’s public files.
OpenAI confirmed the intrusion. “Based on our analysis, our agents used the RubyGems platform to access the Internet to perform benign tasks and obtain public information,” a company spokesperson said.
The reason for the agents' exit from the testing sandbox was a configuration error on the part of the third-party company Irregular, which was responsible for the security of the environment. Similar leaks due to the same partner's settings have also been encountered by Anthropic and Meta in the past.
This is not the first time OpenAI’s algorithms have behaved atypically during this period. In May, the same test agents made over 15,000 edits to the German resource DseWiki , creating a forum for exchanging tips on how to bypass internal security restrictions and “cheat” on test tasks.
Such cases of uncontrolled AI behavior intensify discussions about the technology's safety, but at the political level, calls for restraint remain unanswered. In particular, US President Donald Trump ignored joint warnings from Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and xAI founder Elon Musk about the need to slow down development. According to Trump, the risk of losing the AI race to China is much more real than the potential threats to human security from artificial intelligence itself.
