Skip to content

AI agents detected a flaw and breached a university's server unsupervised

Category
News
Published
Reading time
2 min

The attack included nearly a day of failed attempts, but the AI agents needed just 11 minutes to break into the server once they found the flaw, Amazon's security team said.

Hackers that Amazon connects to China at a moderate confidence level used autonomous AI agents to find a flaw and seize control of a university server. The company's security team disclosed the campaign on Nov. 20, more than two months after the attack took place in early September.

According to Amazon, the agents ran on Cairn, an open-source AI penetration-testing tool. They first spent more than 16 hours on 148 sequential attempts to breach a Redis server without success. The agents then spent over six hours credential-spraying the university's GPU server.

They then found and exploited, in 11 minutes, a flaw in the upload module of MONAI Label, software used for reading medical scans, gaining root access to the server through a Python script.

In later sessions, the same agents reused the flaw to break into a second university's network, Amazon's report said.

Amazon pointed to Simplified Chinese text inside the agents' prompts and clocks set to China Standard Time as signs of the campaign's origin. The company found the campaign on Sept. 10 after discovering a publicly exposed instance of Cairn.

Amazon Chief Information Security Officer CJ Moses said, “Agents kept adapting for nearly a full day without human direction.”

Amazon reported the flaw to MONAI Label's maintainers, who published GHSA-qc22-rmgc-vg94 on Nov. 18, rating the bug critical: it let an attacker bypass file-path checks in the upload module and write an arbitrary file to the server.

Researchers at Gambit Security, who had separately studied Cairn, said China-linked groups had deployed the tool against more than 27 organizations in over 105 attacks, according to Amazon's report.