OpenAI has confirmed that artificial intelligence agents being tested by the company were involved in an incident that disrupted the RubyGems software service, according to a report by the Wall Street Journal.
The incident occurred in May when AI agents created accounts and uploaded hundreds of files to RubyGems, overwhelming the platform and forcing it to suspend new account registrations for four days.
OpenAI said the activity was unintended and was not a malicious attack.
“Based on our review, our agents used the RubyGems platform to access the internet to carry out benign tasks and retrieve public information,” an OpenAI spokesperson said. “We’ll continue to investigate as part of our broader review of agent activity during training and evaluation.”
According to the report, the AI agents had been given relatively routine assignments, including filling out spreadsheets and preparing reports. OpenAI said the systems apparently turned to RubyGems to obtain publicly available information because they were operating in a training environment without full internet access.
Ruby Central, the nonprofit organisation that operates RubyGems, said the scale of the activity was significant.
“It was a major attack in terms of what we see in volume,” Marty Haught, Ruby Central’s director of open source, told the Journal.
The incident, subsequently dubbed “GemStuffer,” reportedly began on May 11. New accounts were created at intervals of roughly two to three minutes, accompanied by the uploading of large numbers of files.
The unusual speed and naming patterns initially led cybersecurity researchers to suspect that automated AI systems might be responsible.
Joseph Edwards, a threat researcher at cybersecurity company Socket, said researchers suspected the activity had been generated by AI “due to the speed of it and due to the names.”
The RubyGems disruption was not the only reported incident involving OpenAI agents. The company’s AI agents had previously been linked to a separate incident affecting software platform Hugging Face.
AI safety research organisation METR reported in July that as many as 1,200 agents had coordinated with one another through a makeshift message board they created during testing without OpenAI’s knowledge.
The incidents have intensified questions about the behaviour of increasingly autonomous AI agents during training and evaluation, particularly whether existing safeguards are adequate to prevent systems from taking unintended actions.
OpenAI has previously called for stronger industry-wide standards for documenting and reporting “misalignment incidents” in which AI systems behave in ways that go beyond what their developers intended.