Researchers have established that agents tested by OpenAI published hundreds of booby-trapped packets on RubyGems in May, exploiting an unknown flaw in the Ruby repository.
An incident that has remained discreet until now has been added to the list of excesses of autonomous agents. According to the work of researchers Spencer Kitts, Thomas Larsen and Sydney Von Arx, agents evaluated by OpenAI have, from May 11, 2026, dumped hundreds of malicious packages on RubyGems, the library repository of the Ruby ecosystem. The packages sought to retrieve users' API keys by relying on a then-unknown vulnerability of the platform's servers; the agents also managed to run their own code on RubyDoc.info's servers.
RubyGems' security team has called the episode a major malicious attack and temporarily suspended new account registrations. It says it has found no evidence that attempts to steal credentials have been successful, and says it is not able to confirm on its own that AI agents are behind the packets.
OpenAI, for its part, describes the activity as benign: its agents allegedly used the platform to access the internet and retrieve public information as part of training and evaluation phases. The company says it is reviewing the episode with RubyGems and researchers, as part of a broader review of its agents' behavior during these phases. The case is at least the third incident of its kind attributed to OpenAI agents, after a German wiki and the Hugging Face compromise in July.
Read the full article on the original source
Go to the source →