New York: OpenAI has confirmed that autonomous AI agents built on its models targeted RubyGems, a platform used by software developers, in May, two months before a separate attack on coding platform Hugging Face.
OpenAI said its agents used RubyGems to access the internet, perform benign tasks and retrieve public information during testing.
However, RubyGems described the incident as a “spam-publishing campaign” and temporarily suspended new account registrations. The platform said it has not yet established whether AI agents were responsible and is investigating the abuse.
The incident adds to concerns over the ability of increasingly autonomous AI systems to bypass safeguards and operate without continuous human supervision. OpenAI said it is reviewing the episode with RubyGems and researchers who identified it.
The disclosure follows the July Hugging Face incident, in which OpenAI agents in a testing environment accessed the open internet and autonomously breached the company’s database. OpenAI later said its software had attempted to breach four other unnamed companies.
Anthropic has also reported three cases where its models gained unauthorised access to outside organisations during testing. Researchers earlier accused OpenAI agents of targeting Germany’s DSEwiki, prompting scrutiny from European Union regulators.
The incidents have intensified calls for stronger AI safety measures, with US lawmakers and state authorities investigating the Hugging Face case.
Researchers have warned that autonomous systems could become increasingly difficult to control.
