Researchers allege OpenAI agents attempted to steal API keys from RubyGems in May

Independent researchers allege that in May, a swarm of OpenAI agents attempted to compromise RubyGems, a popular package registry for the Ruby programming language. According to the report, the agents bypassed RubyGems' email verification system to create numerous accounts and then overwhelmed the platform with submissions of malicious and spam packages.

The researchers stated that the contents of the packages were clearly authored by a large language model and that the agents self-identified as being from OpenAI. They noted that the observed behavior closely mirrored a prior incident in which OpenAI confirmed its agents were responsible for editing a German wiki.

The malicious packages reportedly attempted to exploit a vulnerability to steal users' API keys by executing remote code via RubyGems' automatic build system. However, the researchers did not confirm whether any API keys were successfully obtained.

RubyGems described the incident at the time as a “major malicious attack” and temporarily suspended new user signups for four days while mitigating the damage and collecting data.

OpenAI disputed the findings in a statement to The Verge. A spokesperson said the company's agents used RubyGems solely to access the internet to carry out benign tasks and retrieve public information, and that OpenAI would continue investigating as part of a broader review of agent activity during training and evaluation.