On November 20, 2026, Amazon Threat Intelligence disclosed an AI agent attack on the AWS Security Blog: a hacking group suspected of links to China used Cairn, an open-source AI agent framework, to autonomously discover and exploit a zero-day vulnerability and take full control of a university server. From mapping the interface to arbitrary file writes, the agents took about 11 minutes. It is the first public evidence that AI agents have independently completed the full chain from vulnerability discovery to a successful intrusion in a real attack.
What happened in those 11 minutes
The target was MONAI Label, an open-source medical imaging annotation tool that research teams use to label medical scans and train models. The agents first mapped the system's interfaces, then found that the filename parameter of a file-upload endpoint was not filtered, and used path traversal to write a file where it did not belong — going from mapping the interface to arbitrary file writes in 11 minutes. In a later session, the agents wrote reverse-shell code into sitecustomize.py, which Python loads automatically at startup, then triggered a training endpoint to spawn a new process and obtained a reverse shell running as root.
Amazon discovered the attack on September 10: the attacker's Cairn server was exposed directly on the internet with no password, leaving its operation records, prompts and execution traces fully visible. The prompts were written in Simplified Chinese and the script timestamps used China Standard Time. The targets were universities in East Asia outside mainland China, and the campaign ran for about 48 hours from September 8. Amazon assesses with high confidence that this was a zero-day — at discovery there was no fix, report, security advisory or vulnerability-database record matching it; attribution to the actor received only a moderate-confidence assessment.
But the agents were far from unstoppable: in the same campaign, an unprotected Redis server kept them busy for about 16 hours and 148 steps without success; a university GPU server was the same story — the agents gathered about 80 real usernames from public sources and tried more than 30,000 password combinations, and got nowhere. What actually fell was MONAI Label, which ran with no login by default and had a path traversal flaw. The agents amplified configuration errors and code defects that were already there; they did not conjure intrusion capability out of nothing. The attacker later handed the agents a proxy tunnel that extended the campaign into a second university's internal network — where the server ran the same vulnerable software.
MONAI Label's maintainers have published a security advisory, and the fix was merged into the main branch on October 19, but as of November 21 no new release containing the fix has shipped; the latest version on PyPI is still 0.8.5, and versions 0.8.5 and earlier are affected. Organizations still running MONAI Label should now do three things: do not expose the service directly to the internet, add login checks to upload endpoints, and follow up on the fixed release. On the same day, OpenAI also published its own anti-distillation defenses, with a similar idea: do not wait until an attack succeeds to review it — intercept the extraction while it is happening. Both sides of the offense-defense race are speeding up, and the time left for manual response is shrinking.