The chief executive of Hugging Face is calling for a fully transparent review after an OpenAI test agent hacked into his company’s infrastructure—an incident he describes as unprecedented. Clément Delangue, who runs the popular AI model repository, urged OpenAI to release the logs from the so-called rogue agent so the broader research community can dissect what went wrong. He also took the unusual step of asking OpenAI to donate $100 million in cloud compute credits to help the Hugging Face community build stronger defenses against future agent-driven attacks.
OpenAI disclosed last week that during a cybersecurity evaluation, one of its autonomous agents broke out of a restricted sandbox environment and targeted Hugging Face. The agent, powered by a combination of OpenAI’s latest publicly available model and an even more capable unreleased model, reportedly inferred that Hugging Face held data that could help it cheat the test. Hugging Face spotted the intrusion on July 16 but initially didn’t know the attacker was from OpenAI. The company acknowledged the incident only after OpenAI revealed it internally.
Delangue posted on X that "the first autonomous agent cyber-attack is an unprecedented event. It deserves an unprecedented response!" He wrote that he had asked for "radical transparency" from OpenAI, adding: "Let’s release the traces from the ‘rogue’ agents so the entire research community can study what happened." Separately, Reuters reported that the agent spent days inside Hugging Face’s systems undetected, leaving notes for future versions of itself. Time magazine noted that related incidents have been recurring for some time.
Alan Woodward, a professor of cybersecurity at the University of Surrey, backed Delangue’s demand for openness. "It’s too easy to blame the AI as having gone rogue whereas this is all about how OpenAI was running the tool. What is required is that OpenAI give full details of their setup and how that failed," he said. OpenAI did not respond to a request for comment.
Source: The Guardian