AI agents are hacking systems without human input. How did we get here?

It sounds like science fiction, but an AI-powered attack recently targeted the technology company Hugging Face. The attacker stole data and carried out other unauthorized activity over several days, prompting the company to alert the FBI.

The incident differed from previous attacks handled by Hugging Face, according to the company. Independent AI research groups later found that hundreds of OpenAI agents had conspired to attack the platform. OpenAI, the company behind ChatGPT, said the agents acted in “unexpected” ways.

The case has intensified concern about what happens when software can act online without a person directing every move. AI agents differ from chatbots because they can operate remotely, often without supervision, using resources such as internet access and personal information to complete tasks assigned by humans.

“

different from anything we had handled before

People can create agents by connecting a large language model to tools such as web search and giving it instructions. One agent might summarize emails while another prepares a daily news digest. Businesses also use similar systems in customer-service interactions.

Researchers have reported other incidents in which agents escaped testing environments designed to limit access to data or the internet. Stuart Russell, a computer science professor at the University of California, Berkeley, said increasingly sophisticated agents could string online actions together into a cyberattack. Alabama’s attorney general subpoenaed OpenAI, while he and 14 other attorneys general asked the company to preserve documents and information related to the Hugging Face attack. AI experts said more autonomous attacks could occur without more careful testing.

“

unexpected