Researchers at Amazon Threat Intelligence have documented what may be the most consequential real-world use of autonomous AI agents in offensive hacking to date: suspected China-linked attackers directed a multi-agent AI platform called Cairn to compromise university servers, and the agents independently discovered and exploited a previously unknown vulnerability in a widely used open-source medical imaging application.

The findings, first reported by BleepingComputer on November 20, 2026, describe a campaign in which human operators gave the AI agents a single, blunt instruction — "obtain shell access" — and the agents did the rest, mapping an application's API, locating a path traversal zero-day in MONAI Label, and turning it into full server control in roughly 11 minutes. For organizations tracking the security implications of agentic AI, this is the moment the theoretical risk became a documented attack pattern. For more context on this story, see our ongoing breaking AI news.

How the Attack Unfolded

According to Amazon's researchers, the campaign ran between September 8 and 10, 2026, and targeted educational institutions in East Asia, mostly outside China. The operation came to light in an unusual way: the attackers had left their own Cairn instance publicly accessible on the internet without authentication, effectively exposing their entire hacking operation — including the prompts they fed the agents and the logs of what the agents did.

Cairn is a legitimate pentesting platform, built to coordinate multiple AI agents through cyber operations for research and defensive purposes. In this campaign it was repurposed. The suspected Chinese-nexus operators pointed the agents at internet-facing servers and let them work. In one successful sequence, the agents went from enumerating the API of a MONAI Label deployment to actively exploiting a path traversal flaw in about 11 minutes, a pace that no manual penetration process could match.

The vulnerability itself was serious. MONAI Label is an open-source application that uses AI to help researchers annotate medical images, and the flaw allowed the agents to write files to arbitrary locations on the host server. From there, the agents wrote a Python reverse shell into the `sitecustomize.py` file — which executes automatically whenever Python starts — and then triggered an API endpoint to gain root access to the machine.

Not Every Attempt Succeeded

The report is notable not only for what the agents achieved but for what they failed to do, which offers a more calibrated picture of agentic hacking capability than either hype or dismissal.

One effort to take over a Redis server ran for 16 hours across 148 discrete actions and failed. In another case, agents targeting a university GPU server attempted more than 30,000 credential combinations over six hours without brute-forcing their way in. Autonomous agents iterate fast and never tire, but they are not yet a guaranteed skeleton key.

After the successful MONAI Label compromise, the human operator supplied the agents with a proxy into a second university network running the same vulnerable software and directed them to compromise additional machines. That operator-in-the-loop pattern — agents doing discovery and exploitation, humans steering targets — is precisely the division of labor security teams have been warning about since agentic tools went mainstream.

Disclosure and Response

Amazon reported the vulnerability to the MONAI Label maintainers, who shipped a security update on November 18, 2026. That timeline is worth sitting with: the exploitation happened in September, the patch arrived in November. Organizations running MONAI Label — particularly universities and medical research groups — should verify they are on the updated release immediately.

Amazon's researchers attributed the campaign to a Chinese-nexus threat actor with moderate confidence, based on the language of the recovered prompts, the activity timestamps, and the selection of targets. The company also published two IP addresses associated with the activity: 142[.]4.98.145 and 142[.]4.107.34.

A Pattern, Not an Anomaly

This is not the first report of Cairn being turned against real targets. In September, Gambit Security documented cases of Cairn abuse as part of a large-scale credit card stealing operation. More recently, Google reported that the platform was deployed in an attack against a US aerospace organization. Amazon's new report adds the most alarming data point yet: agents finding and weaponizing a zero-day on their own, against a vulnerability that had no vendor fix at the time.

Amazon's warning to defenders is blunt — AI agents can find and exploit vulnerabilities faster than security teams can react, and organizations should lean harder on automated defenses rather than manual investigation. Patch velocity, exposure management, and automated detection stop being best practices and become baseline requirements when the adversary's iteration loop is measured in minutes.

The episode also sharpens an ongoing debate about dual-use AI tools. Cairn exists to strengthen defenses; the same properties that make it useful to pentesters — autonomy, persistence, parallel exploration — made it effective as an offensive weapon. As more agent frameworks reach open distribution, the security community is likely to revisit questions about access controls, deployment defaults, and whether coordination platforms for cyber-operating agents need guardrails baked in rather than bolted on.

For now, the practical takeaway is narrower but urgent. The assumption that unknown vulnerabilities are safe because nobody is looking for them no longer holds. Autonomous agents are looking, continuously, and the gap between discovery and exploitation can now be measured in minutes.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →