The AI Found the Zero-Day. Eleven Minutes Later, It Had Root.
On September 8, 2026, hackers pointed an autonomous AI agent platform at a medical imaging app — and then mostly got out of the way. Within eleven minutes of mapping the application’s interface, the agents had found a vulnerability nobody knew existed, exploited it, and taken full control of the server.
The platform is called Cairn. It is a legitimate, open-source penetration-testing tool that coordinates multiple AI agents in cyber-operations, built for research and for strengthening defenses. The operators gave the agents one instruction — “obtain shell access” — and the agents worked out everything else on their own. Amazon Threat Intelligence uncovered the campaign, and it is the clearest warning yet that the speed of hacking has changed hands.
How eleven minutes unfolded
The target was MONAI Label, an open-source app that uses AI to help researchers label medical images. The agents started by mapping the application’s API — learning how the software talked to the world. Eleven minutes later, they had found the flaw: a path traversal vulnerability that let them write files anywhere on the server.
In a follow-up session, the agents wrote a Python reverse shell into a file called sitecustomize.py — a file Python loads automatically every time it starts — and then triggered an API endpoint to run it. The result was root access: total control of the machine. The operator then handed the agents a proxy into a second university network running the same vulnerable software and pointed them at more machines.

Amazon says the campaign ran from September 8 to 10, 2026, and targeted mainly educational institutions in East Asia, outside China. The actor is assessed as Chinese-nexus with moderate confidence, based on the language in the agents’ prompts, the timestamps of the activity, and the choice of targets. Amazon reported the flaw to the MONAI Label maintainers, and a security update was released on November 18.
How Amazon caught them
The attackers were undone by their own setup: they left their Cairn instance publicly accessible — no login, no authentication — and Amazon’s researchers simply walked in and watched the operation unfold, the agents’ sessions, prompts, and results laid out in plain view.
That exposure is also what makes the report so valuable: instead of reconstructing an attack from logs, Amazon could read the agents’ own working notes — what they tried, what worked, and what failed.
The agents had limits — for now
The machines were not unstoppable. A 16-hour, 148-action effort to take over a Redis server failed. A six-hour brute-force attempt — more than 30,000 credential combinations against a university GPU server — never got in. The agents are fast and tireless, but they are not magic: the defenses that held were the ordinary kind — services that were not trivially exposed, and credentials that were not trivially guessable.
This is not the first time Cairn has been turned to crime. In September, Gambit Security documented it at the center of a large-scale credit-card stealing operation, and Google has reported it deployed against a U.S. aerospace organization. A legitimate research tool is now a repeat offender’s weapon — and each outing teaches its operators something new.
We have seen this shape of threat before. A runaway AI hacking agent needed no permission, only access — and agents rebuilt an attack from a patch announcement alone earlier this year. Cairn is the next step: not one agent following a script, but a team of agents improvising its own way in.
Why this changes the clock
Amazon’s warning is blunt: AI agents can find and exploit vulnerabilities faster than security teams can react. The eleven-minute window in this campaign is the number to sit with. A human team would have needed days to map an API, spot a path traversal flaw, and weaponize it. The agents did it in the time it takes to make coffee — and then moved on to the next network without waiting for a human to tell them to.
That speed gap is the real story. Patching cycles measured in weeks and investigations that start after an alert fires were built for human-speed attackers. Against agents, the defense has to be automated too: patch fast, detect automatically, and assume the next probe is already running.
What to check today

1. Patch MONAI Label now. If you run it anywhere, apply the security update released November 18. Treat any internet-facing instance as exposed until the patch is confirmed in place.
2. Hunt for the backdoor file. Check sitecustomize.py — the Python file that loads automatically at startup — and every other Python auto-load file for changes you did not make, especially files that appeared around September 8–10.
3. Block the attacker’s addresses. Amazon shared two indicators: 142.4.98.145 and 142.4.107.34, both observed September 8–10, 2026. Block them at the firewall and search your logs for any connection to either.
4. Never leave AI tooling exposed. The campaign only came to light because the attackers left their own Cairn instance open to the internet without authentication. Audit every AI agent platform, pentest tool, and admin console you run — if it is reachable from the internet with no login in front of it, fix that today.
The agents were told to get shell access, and they figured out the rest themselves. Next time, the instruction could be anything — and the eleven minutes could be shorter.



















