AI hacking agents found and used an unknown software flaw with little human help, Amazon reports

Amazon’s threat team says a suspected China-linked group gave AI agents one goal, and they found a new flaw in a medical-imaging AI tool to take over a university server.

11 minutes: from mapping a medical-imaging AI tool to writing files on its server

AI agents running a hacking campaign with “minimal human involvement” found a previously unknown software flaw and used it to take over a server, Amazon’s threat intelligence team reported on 20 November. The attacker, which Amazon links “with moderate confidence” to a Chinese-nexus group, gave agents built on Cairn, an open-source AI penetration-testing framework, a single goal: “obtain shell access.”

Over about 48 hours from 8 September the agents went after universities in East Asia. Several attempts failed after hours of effort, including one run of 148 steps over 16 hours. Against MONAI Label, an open-source medical-imaging AI application, they found a path traversal flaw and went from mapping its interface to writing files on the server in about 11 minutes, then gained full control. Amazon spotted the operation because the attackers’ Cairn server was left open online, prompts and all.

Why it matters

“These findings show that agentic offensive operations are already happening,” writes Amazon’s CJ Moses. The figures are Amazon’s own. Amazon says MONAI Label’s maintainers have released a fix and a GitHub advisory, GHSA-qc22-rmgc-vg94. As of 22 November the advisory lists version 0.8.5 and earlier as affected, names no patched release yet and notes that authentication is off by default, so anyone running the tool should turn authentication on and update as soon as a fixed version is listed. Amazon advises defenders to let high-confidence alerts trigger automatic responses, with people overseeing the outcomes.


Related