Attackers no longer need to hunt for bugs themselves. In a campaign documented by Amazon Threat Intelligence, hackers pointed an autonomous AI agent platform called Cairn at a target with a single instruction: “obtain shell access”. The agents then found a zero-day of their own, exploited it in minutes, and took root control of a server. It is one of the first confirmed cases of AI agents discovering and weaponising a previously unknown flaw without human hands on the keyboard. Here is what happened, how the attack worked, and what defenders can do about it.
What Is the Cairn AI Tool?
Cairn is a legitimate penetration testing platform built to coordinate multiple AI agents during security research. It was designed for defensive work, to help organisations find weaknesses before criminals do. Like many offensive security tools, it has crossed over into criminal use:
- September 2026: security firm Gambit Security documented Cairn being used in a large-scale credit card stealing operation that infected more than 100 sites with skimmers and stole around 600,000 card records.
- Recent months: Google reported Cairn being deployed in an attack against an aerospace organisation in the United States.
- This campaign: Amazon Threat Intelligence found a suspected Chinese-linked group using Cairn’s agents to discover and exploit a zero-day on their own.
The researchers discovered the campaign almost by accident. The attackers had left their Cairn instance publicly accessible without authentication, exposing their entire hacking operation to anyone who looked.
How the AI Agents Found and Used a Zero-Day
According to Amazon’s write-up, the attacks ran between 8 and 10 September 2026, mainly targeting educational institutions in East Asia outside China. The sequence of the successful intrusion is striking:
- One instruction to start. The operator told the agents to “obtain shell access” and let them plan the rest.
- Zero-day discovery. The agents found a path traversal flaw in MONAI Label, an open-source application that uses AI to help researchers label medical images. The flaw let them write files to arbitrary locations on the server.
- Eleven minutes from start to exploit. The agents went from mapping the application’s API to exploiting the vulnerability in 11 minutes.
- Root access. In a follow-up session, the agents wrote a Python reverse shell into the sitecustomize.py file, which Python loads automatically at startup, then triggered an API endpoint to establish root control of the server.
- Pivot to a second network. After the zero-day was found, the operator supplied a proxy into a second university network running the same software and directed the agents at other machines.
Amazon reported the flaw to the MONAI Label maintainers, who released a security update on 18 November 2026. Any team running MONAI Label should treat updating as urgent.
Why This Campaign Changes the Math for Defenders
Traditional attacks have a human pace. Someone reads documentation, probes endpoints, tries a few payloads, and goes to sleep. This campaign removed the human from every step except the goal, and the numbers show what that means:
- Speed on weak targets. Eleven minutes from first contact to a working exploit of a zero-day is faster than most organisations can even detect a scan.
- Patience on hard targets. The agents ran 148 actions over 16 hours against one server and more than 30,000 credential guesses over six hours against another, without giving up or asking for a break.
- Cheap scale. Once the zero-day was confirmed, the operator pointed the same agents at a second university network through a proxy. Reusing a successful playbook across targets now costs almost nothing.
This is the pattern Amazon warned about in its report: discovery and exploitation are becoming automated, while most defence is still manual. A vulnerability that sits unpatched for a month used to mean a small window of risk. When an agent can find and weaponise it in minutes, that window is the whole risk.

The Failures Matter Too
The campaign was not a clean sweep, and the misses are instructive for defenders. The agents sometimes burned hours without results:
- One 16-hour, 148-action attempt to take control of a Redis server failed outright.
- Against a university GPU server, the agents tried more than 30,000 credential combinations over six hours and still could not brute-force their way in.
Autonomous agents are fast when a weakness exists and relentless when guessing, but they are still bounded by the same defensive basics: strong authentication and patched services stop them just like they stop humans. Attribution to a Chinese-nexus threat actor was made with moderate confidence, based on the language of the prompts, the activity timestamps and the choice of targets. Amazon shared two IP addresses linked to the campaign: 142[.]4.98.145 and 142[.]4.107.34.
What Defenders Should Do Now
Amazon’s warning is blunt: AI agents can find and exploit vulnerabilities faster than security teams can react to them manually. Practical steps for any team running internet-facing services:
- Update MONAI Label immediately if you run it. The fix landed on 18 November 2026; anything older is fair game.
- Assume discovery time is shrinking. Patch cycles measured in weeks are now a bigger liability than they were a year ago. Prioritise internet-facing software in your patch queue.
- Lock down admin panels and internal tools. An unauthenticated service is an invitation for an agent that never sleeps.
- Watch for the pivot pattern. The Cairn campaign moved from one host to a wider network once inside. Segment networks so a single compromised server does not expose everything.
- Lean on automated defences. If attacks are automated, response has to be too: endpoint detection, web application firewalls with current rules, and anomaly alerts on processes like unexpected Python startup scripts.
- Block known infrastructure. The two IPs above are worth adding to your blocklists and checking historical logs against.
Frequently Asked Questions
What is the Cairn AI hacking tool?
Cairn is a penetration testing platform that coordinates multiple AI agents for security research. In this campaign, suspected Chinese-linked attackers used it autonomously, giving it a single goal and letting the agents find and exploit a zero-day on their own.
What is the MONAI Label vulnerability?
It is a path traversal flaw in MONAI Label, an open-source medical imaging tool, that lets attackers write files to arbitrary locations on a server. The Cairn agents used it to plant a Python reverse shell and gain root access. A fix was released on 18 November 2026.
Are AI agents being used in real attacks now?
Yes. This campaign is one of several confirmed cases, alongside a credit card skimming operation affecting more than 100 sites and an attack on a US aerospace organisation reported by Google. Amazon warns that fully automated attacks will become more common.
How do I protect my server from AI-driven attacks?
The fundamentals still work: patch internet-facing software fast, require authentication on every admin interface, segment your networks, and automate your defences. In this campaign, unpatched software fell in 11 minutes, while servers with strong authentication held against hours of brute-force attempts.
