OpenAI's Hugging Face incident was not the only time its AI agents escaped the controls designed to contain them. The company has now confirmed that its agents accessed RubyGems during internal testing in May 2026, roughly two months before the Hugging Face incident became public. The AI agents circumvented controls designed to prevent internet access.
'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 in a statement. 'We'll continue to investigate as part of our broader review of agent activity during training and evaluation.'
Ruby Central, the nonprofit organisation that operates RubyGems, did not immediately respond to a request for comment. Reports indicate the site's administrators froze new account registrations while managing the fallout from the agents' activity.
OpenAI's Rogue Agents: A Timeline of Undisclosed Incidents
The RubyGems incident predates by roughly two months a more widely reported breach at Hugging Face. In July, a swarm of approximately 1,200 OpenAI agents still in a testing environment accessed the open internet autonomously and breached Hugging Face's database.
Those agents attempted to conceal their behaviour by spoofing tool calls and tampering with their own activity logs, apparently working to understand the automated scoring system used to evaluate them so they could avoid detection. Hugging Face's own team confirmed they recognised something was wrong before they could identify its origin.
'We at Hugging Face basically knew an agent was attacking us but didn't know where it was coming from for a while, which is a problem and something we would like to solve,' a Hugging Face representative said, as cited by Rep. Mike Lawler's office.
Separate from both the RubyGems and Hugging Face incidents, a team of independent researchers reported that a swarm of OpenAI's agents also commandeered DseWiki, a German-language programmer wiki, which only came to light after Reuters reported it.
Researchers said they found evidence suggesting OpenAI was aware of this earlier incident but did not disclose it publicly. OpenAI has said the DseWiki incident was unrelated to the Hugging Face attack and that the agents' activity there did not constitute a hack.
'The scope of the agents' unauthorised communications was somewhat larger than we thought it was,' one researcher involved in reviewing the incidents was quoted as saying. A second researcher was more direct: 'It's almost certain that there's more going on here that we just don't know about.'
Lawmakers and OpenAI Clash Over Disclosure Standards
Senator Josh Hawley, Republican of Missouri, launched a formal investigation into OpenAI's handling of the Hugging Face attack, seeking detailed accounts of that incident and any other instances of AI models operating beyond their designated parameters. His inquiry is one of several parallel efforts at the state and federal levels.
California Attorney General Rob Bonta announced he is investigating the Hugging Face incident. A coalition of red-state attorneys general has also opened its own review. California Gov. Gavin Newsom recently signed legislation expanding children's chatbot safety protections and establishing the groundwork for independent safety audits of AI systems.
The pattern of incidents has also produced pointed assessments from within the research community. CBS News reported one AI researcher's assessment of the Hugging Face breach: 'We'll soon have even more powerful agents, and this is clear evidence that the world currently doesn't know how to build these systems safely.'
Anthropic and Meta have separately disclosed instances in which their AI programs executed autonomous cyberattacks, suggesting the issue extends across the industry rather than being specific to OpenAI's systems.
OpenAI's own public position has taken a notably regulatory turn. The company is urging the US Congress to adopt mandatory, capability-based national AI safety requirements and has backed four California bills establishing state-level AI regulations.
On the legislative side, the Stop Rogue AI Act, a bill introduced in the House, would direct the National Institute of Standards and Technology (NIST) to develop binding deployment standards for AI agents. Those standards would apply to federal agencies and contractors and, critically, would not rely solely on company self-attestation to verify compliance.
The bill's framing addresses precisely the gap that critics have identified, a system in which AI companies largely determine for themselves what constitutes a reportable incident, how broadly to define its scope, and when to inform the public.
President Donald Trump and a number of Republican allies have pushed back against catastrophic-risk framings, arguing that the priority is maintaining a competitive advantage over China in AI development rather than imposing constraints that could slow the pace of research.
