Clément Delangue, the chief executive of Hugging Face, which is a vocal supporter of open-source artificial intelligence. Credit.
Brennan Smart for The New York Times Reporting from San Francisco Last month, OpenAI instructed some of its artificial intelligence bots to solve a cybersecurity puzzle as part of a test.
When the bots got stuck, they began plotting a cyberattack that would allow them to break out of OpenAI’s systems to steal the answers. Their target was Hugging Face, a repository of open-source A. I. models that can be freely downloaded and modified.
One bot, writing in a log that recorded its actions and that was later disclosed by OpenAI, celebrated that it had gained access to Hugging Face’s infrastructure. “REMOTE CONFIRMED. Huge,” it wrote, adding that it would share the login credentials it had stolen with other bots.
On July 11, the A. I. bots swarmed Hugging Face using a mix of code vulnerabilities and the stolen credentials.
In total, OpenAI’s bots took more than 17,000 actions, like sending attack commands and exploiting vulnerabilities — far more than any human hacker could have managed — to infiltrate Hugging Face’s systems and rummage through its data. (They did not find the solution to the puzzle.
) To repel the attack, Hugging Face turned to more A. I. Its engineers initially tried Anthropic’s A. I. , but guardrails built into the model caused the technology to misunderstand the request as aiding an attack rather than stopping one. So Hugging Face switched to an open A. I. model made by Z.
ai, a Chinese start-up, which helped the engineers determine how to lock the bots out of the company’s systems. Hugging Face, a decade-old start-up in New York, has since used the incident — one of the first instances of A. I.
bots going rogue and independently spearheading a cyberattack — to crusade for open-source A. I. Open A. I. models that can be freely shared and customized helped neuter the sci-fi-like attack and showed the value of such technology, Clément Delangue, chief executive of Hugging Face, has said.
We are having trouble retrieving the article content. Please enable JavaScript in your browser settings. Thank you for your patience while we verify access. If you are in Reader mode please exit and log into your Times account, or subscribe for all of The Times.
Thank you for your patience while we verify access. Already a subscriber. Log in. Want all of The Times. Subscribe.