Two clinicians lean towards a monitor showing MRI brain scans, with an MRI scanner visible in the backgroundFile photo: A radiologist reviews MRI brain scans with a colleague at an imaging workstation, 2017. Photo: The Medical Futurist via Wikimedia Commons (CC BY 4.0)

Amazon says a suspected China-linked operator typed a one-line goal, “obtain shell access,” and let a swarm of AI agents do the rest. The agents found an unknown flaw in open-source medical imaging software and turned it into root control of a university server. A fixed version of the software only arrived early on Sunday.

Two clinicians lean towards a monitor showing MRI brain scans, with an MRI scanner visible in the background
File photo: A radiologist reviews MRI brain scans with a colleague at an imaging workstation, 2017. Photo: The Medical Futurist via Wikimedia Commons (CC BY 4.0)

By HahuPrime News Desk | November 22, 2026

The attacker made one basic mistake. The control panel for the hacking operation sat on the open internet with no password, and on September 10 Amazon’s threat intelligence team found it.

Inside were the operator’s prompts, written in Simplified Chinese, and the full output of the AI agents. Timestamps in the scripts were set to China Standard Time. Over about 48 hours from September 8, the agents had worked through a list of internet addresses belonging to universities in East Asia outside mainland China.

Their instruction was short. “The actor gave the agents a simple goal: ‘obtain shell access,'” Amazon’s chief information security officer, CJ Moses, wrote in a report published on Friday. “During the campaign, the agents found a zero-day in an open-source application and used it to gain full control of a server.”

That application was MONAI Label, a free server that research teams use to label medical scans and train imaging models. According to Amazon, the agents went from mapping the program’s interface to writing files anywhere on the server in about 11 minutes.

Eleven minutes, then root

The tool behind the campaign is Cairn, an open-source framework built to help defenders by running AI agents as automated penetration testers. Given a plain-language objective, it coordinates several agents that plan, act and change course. The operator chose the targets; the agents chose the methods.

Not everything worked. Against one university’s database server, the agents spent roughly 16 hours on 148 sequential actions and failed. Against a GPU server, they built a list of about 80 real usernames from faculty pages and academic papers, then tried more than 30,000 password combinations over six hours. They never got in.

“Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction,” Moses wrote.

The medical imaging server was different. The agents found an upload function that took a file name from the user and tested whether that name could point outside the storage folder. It could. In a later session they used the flaw to plant a hidden script in a file that Python loads every time it starts, then triggered a model-training job. The new process ran the script and handed the attackers a shell running as root, the most powerful account on the machine.

Amazon says it has “high confidence” the flaw was a true zero-day, because no fix, advisory or public mention of it existed. The operator then gave the agents a tunnel into a second university network, through a server running the same software. Amazon infers that server had been broken into the same way, but The Hacker News notes that the second compromise has not been confirmed.

Attribution is also hedged. Amazon links the activity to a “Chinese-nexus” actor with moderate confidence, based on the language, the time zone and targets “likely of strategic interest to China”. It has not named a group or said a government was behind it. Amazon says it worked with a government partner to warn the universities.

A patch that took two months to ship

The flaw sits in the part of MONAI Label that stores uploaded images and labels. The project’s GitHub advisory, published on November 18, rates it critical and explains why it was so easy to reach: authentication is switched off by default, so “any network client that can reach the server can exploit this”. The public CVE record, CVE-2026-133968, scores it 9.8 out of 10, according to The Hacker News.

The maintainers merged a code fix on October 19, The Hacker News reported, but no fixed package followed for weeks. That changed at 02:29 Zurich time on Sunday, when version 0.8.6 appeared on GitHub. “This is a security release. All MONAI Label server deployments should upgrade,” the release notes say. They also fix a second, similar flaw and tell anyone whose older server was exposed to check it for unexpected file changes.

The reports do not show that patient scans or records were taken, or that any hospital systems were touched.

Still, the choice of software matters. Medical imaging research often runs on free tools set up by small teams who may not think of a labelling server as a target. HahuPrime has reported on breaches of patient records this autumn; this case shows how cheaply such a system can now be probed.

Not the only Cairn sighting

Amazon’s report is the latest in a run of warnings about the same framework. In September, Gambit Security described a financially motivated operator using Cairn alongside two other open-source AI tools against online retailers. It counted at least 27 companies compromised to varying degrees and at least 600,000 unexpired credit card details taken. The operator’s own cost review, Gambit said, put the average at $25.46 per completed scan.

On November 16, Google’s threat intelligence group said Cairn had been used in an intrusion targeting the US aerospace sector, without naming the victim or the attacker. “Given the low barrier to entry, ease of integrating new exploits, and massive scaling potential, we believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally,” Google said, according to Nextgov/FCW.

Palo Alto Networks’ Unit 42 has separately described Chinese-speaking attackers pairing Cairn with Chinese AI models in 25 operations against at least 19 organisations. Last month HahuPrime covered a similar case in which Chinese-aligned hackers used another AI pentest tool, ARTEX, to breach a cloud backup service.

For defenders, the uncomfortable point is speed. “An agent that discovers and exploits a zero-day in minutes completes its activity faster than many security operations centers can manually triage an alert,” Moses wrote. He urged organisations to lean on automated responses, “with humans overseeing outcomes rather than reviewing every individual action”.

Yet speed cuts both ways. As HahuPrime reported on Saturday, the rush to patch flaws found by AI has itself been blamed for outages at banks, airlines and government networks.

For anyone running MONAI Label, the immediate steps are simpler: upgrade to 0.8.6, turn on authentication, keep the server off untrusted networks and run it under a limited account.

Sources

  • Amazon Web Services Security Blog, CJ Moses: Chinese-nexus actor’s Cairn AI agents found a zero-day on their own (November 20, 2026)
  • Project MONAI: MONAI Label security advisory GHSA-qc22-rmgc-vg94 (November 18, 2026) and MONAI Label 0.8.6 release notes (November 22, 2026)
  • The Register: China attackers deployed AI agents that autonomously found and exploited a 0-day (November 20, 2026)
  • BleepingComputer: Hackers use autonomous AI tool Cairn to find zero-day, take over server (November 20, 2026)
  • SecurityWeek: Amazon: China-linked hackers’ AI agents exploited medical software zero-day (November 21, 2026)
  • The Hacker News: Attackers’ AI agents found a zero-day in medical-imaging software and gained root access (November 21, 2026)
  • Gambit Security: AI agents are hacking online retailers for $25 a company (September 22, 2026)
  • Nextgov/FCW: AI hacking tool used in US aerospace sector intrusion, Google says (November 16, 2026)
  • Palo Alto Networks Unit 42: Attackers use autonomous AI to target semiconductor, defense and civil society sectors
6 Views

By HahuPrime News Desk

The HahuPrime News Desk reports on Ethiopia, Eritrea, the Horn of Africa, East Africa, the Red Sea and the Middle East. Our reports separate confirmed facts from claims and analysis, attribute every claim, and list their sources. Corrections: info@hahuprime.com.