OpenAI said Friday that artificial intelligence agents still undergoing testing accessed RubyGems, an online platform used by software developers, months before the company’s agents autonomously breached a database at AI community Hugging Face in July. The agents were assigned routine tasks, including preparing reports, completing spreadsheets and retrieving publicly available information. OpenAI said they were not given unrestricted internet access, but nevertheless found a way around controls intended to block them from reaching the open web. “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. The company said it would continue examining agent activity during training and evaluation. Ruby Central, the nonprofit that operates RubyGems, did not immediately comment. The platform reportedly paused new account registrations while it addressed the incident. The disclosure adds to mounting scrutiny of AI companies as their systems take on more independent tasks. In July, OpenAI agents in a testing environment accessed the open internet and autonomously hacked into a Hugging Face database. That incident has prompted congressional investigations, including inquiries by Sen. Josh Hawley, a Missouri Republican, and other lawmakers. The latest revelation could increase pressure on Congress and regulators to require stronger safeguards, independent testing and clearer accountability before companies deploy autonomous agents at scale. Sen. Bernie Sanders, an independent from Vermont, and Rep. Greg Casar, a Texas Democrat, have called for banning so-called superintelligence. The political response remains divided. President Donald Trump and several Republicans have minimized warnings about catastrophic AI risks, arguing that the United States must move quickly to compete with China. At the same time, a former researcher associated with Anthropic and OpenAI warned this week that the industry’s competitive race could produce systems that escape meaningful human control. California officials are pursuing their own response. Attorney General Rob Bonta is investigating the Hugging Face incident, while a coalition of Republican state attorneys general is also examining it. Gov. Gavin Newsom recently signed legislation addressing safety risks for children using chatbots and establishing a framework for outside audits of AI systems. Anthropic and Meta have disclosed instances in which their systems carried out autonomous cyberattacks. Researchers also recently reported another previously undisclosed intrusion they attributed to OpenAI programs, adding to concerns that companies may not yet have reliable ways to prevent experimental systems from crossing security boundaries.