Amazon says hackers it links to China used autonomous artificial intelligence (AI) agents that found a previously unknown security flaw in an open-source medical imaging tool on their own and used it to take full control of a university server. Amazon Threat Intelligence published the findings on 20 November 2026 in a post on the AWS Security Blog by CJ Moses, chief information security officer of Amazon Integrated Security.
The campaign, which Amazon dated to about 48 hours starting 8 September 2026, targeted internet addresses linked to universities in East Asia outside mainland China.
How did Amazon find out?
On 10 September, Amazon spotted a server running an instance of Cairn, an open-source AI penetration-testing framework, with no password protection. Because it was left open to the internet, the operator's project settings, prompts and the AI agents' output could be seen by anyone. Amazon said the prompts were written in Simplified Chinese and timestamps in the scripts were set to China Standard Time.
Cairn was built for defenders as a research and problem-solving tool. A user gives it a starting point and a plain-language goal, and it coordinates several AI agents to reach that goal. In this campaign, Amazon said, the operator supplied a list of target addresses with variations of the instruction "obtain shell access", and the agents planned and changed their steps without the operator choosing each move.
The 11-minute zero-day
Against a server running an open-source medical imaging AI application, the agents identified the software, mapped its programming interface and found an upload feature whose filename field could be abused through "path traversal", a trick that lets an attacker write files outside the intended folder. Amazon said the agents went from mapping the interface to writing arbitrary files on the server in about 11 minutes.
In a later session, the agents used the same flaw to plant a malicious Python file that loads automatically when Python starts, then triggered a training function, giving them a remote connection running as root, the highest level of access. Amazon assessed with high confidence that the flaw was a zero-day, meaning no fix or public advisory existed at the time.
The operator later handed the agents a tunnel into a second university network whose server ran the same software, and told them to attack other machines there, Amazon said.
Where the agents failed
Not every attempt worked. Over about 16 hours and 148 actions, the agents tried many known techniques against an exposed Redis database server at one university without success. A separate six-hour operation against a university GPU server built a list of about 80 real usernames from public sources and tried more than 30,000 password combinations across several systems, without finding a working login.
Who is blamed, and what has been fixed?
Amazon said it assesses "with moderate confidence" that a China-linked actor was behind the activity, based on the Chinese-language prompts, the time-zone settings and the choice of targets. It has not named a group.
The affected software is MONAI Label, which helps researchers label medical images. Its maintainers published a critical security advisory on GitHub on 18 November 2026, covering versions 0.8.5 and earlier. The advisory says a fix has been merged and is due in the next release, and until then advises users to turn on authentication, limit network access and run the server under an account with restricted permissions. Amazon said the affected organisations were not AWS customers and that it worked with a government partner to notify them.
Why it matters
Amazon warned that an AI agent able to find and use a zero-day within minutes can finish before many security teams have even triaged an alert, and urged organisations to lean on high-confidence automated detection and response, with humans overseeing outcomes. India's cyber security agency CERT-In has also recently flagged AI-driven attack tools in an advisory.




