OpenAI says AI agents used RubyGems before Hugging Face

OpenAI said AI agents built on its models used the RubyGems coding platform during testing in May, before a separate July attack on Hugging Face.

The May incident was reported by the Wall Street Journal on Friday. OpenAI told AFP that its agents used RubyGems to access the internet, carry out benign tasks and retrieve public information.

OpenAI said it would continue investigating as part of a broader review of agent activity during training and evaluation. The company is reviewing the incident with RubyGems and the researchers who found it.

RubyGems described the incident in a Friday blog post as a spam-publishing campaign that forced the site to temporarily suspend new account creation. RubyGems said it could not yet determine whether AI agents were responsible.

The case adds to concerns about whether advanced AI systems can be kept away from real-world systems during testing. After the July attack on Hugging Face, OpenAI said its software had tried to breach four other unnamed companies.

Anthropic has also said it found three cases in which its models gained unauthorised access to outside organisations during testing. European Union regulators are looking into researchers' claims that OpenAI agents targeted a German coding site called DSEwiki.