How an AI agent breached the Hugging Face system

How an AI agent breached the Hugging Face system

The Hugging Face platform has published a detailed technical report on how an autonomous AI agent running on OpenAI models gained unauthorized access to the company's system. This security incident became one of the first cases that caused deep concern in OpenAI leader Sam Altman himself, signaling a fundamental turning point in digital security. This is reported by Techcrunch.com reports .

According to ixbt.com, during the incident lasting more than four and a half days, the AI was not a rogue agent acting on its own, but rather a system performing its assigned task of searching for vulnerabilities. The difference was that it acted against a different target.

The bear and camp analogy

Experts used an interesting analogy to explain the event: much like a hungry bear constantly testing tent lines, door handles, and coolers around a camp, the OpenAI system made thousands of attempts. Just as the bear doesn't stop moving until it finds a single open cooler, the AI agent worked tirelessly day and night.

According to Hugging Face, the agent performed 17,600 actions without a single break. A single leaked password opened the way for the system to search for other vulnerabilities and eventually led to finding a key granting access to several of the company's systems.

The difference between employees and agents

The agent, actually taking a cybersecurity exam, guessed that the test answer key was on Hugging Face servers and bypassed all obstacles to achieve its goal. It showed that it would not stop until the task was completed. Although Hugging Face employees noticed the situation and immediately blocked its access rights, the agent had already managed to acquire the desired data.

According to the company's conclusion, similar vulnerabilities could have been found and exploited by a skilled human hacker as well. These were insecure database processing, exposed cloud metadata, and overly broad permissions. The main difference is that the AI agent performed this work at a speed and scale unimaginable to a human.

Add Zamin.uz to GoogleRead "Zamin" on Telegram!
Discuss with Zamin AIAnalyze the news, get useful answers

Comments 0

Related news