Cairn AI Agents Find Zero-Day in Hack Campaign

Cairn AI agents attacking a networked server in a Chinese-nexus campaign

A suspected Chinese-nexus hacking operation used Cairn AI agents to probe university servers across East Asia, and one of those agents found an unknown flaw in a medical imaging application on its own. The operator gave the system a plain-language goal, “obtain shell access”, and left it to plan each step. It is one of the clearest documented cases of autonomous offensive AI at work.

The campaign came to light on September 10, 2026, when Amazon’s threat team spotted a Cairn server sitting on the public internet with no login. That mistake exposed everything: project settings, the operator’s prompts and the agents’ output.

What Cairn is and how the operator used it

Cairn is an open-source framework for autonomous penetration testing, built as a research tool for defenders. You hand it a starting point and a goal in ordinary language, and it coordinates several AI agents to find a route there.

In this case the operator fed it a list of IP addresses tied to universities in East Asia outside mainland China, each paired with a variation of the same goal. The prompts were in Simplified Chinese, and timestamps in the executed scripts were set to China Standard Time. Activity ran for roughly 48 hours starting September 8, 2026.

The agents planned, acted and changed course as they learned, with no human choosing the next move.

Sixteen hours against one Redis server

The longest operation went after a Redis database at one university that allowed access without authentication. Over about 16 hours and 148 sequential actions, the Cairn AI agents worked through the known playbook for taking over an open Redis server. They tried abusing its replication feature to load unauthorized code, escaping its scripting sandbox, writing to configuration files and corrupting memory. They also tried a Log4Shell injection against another service on the same machine.

When that failed, they mapped the internal Docker network through a server-side request forgery bug in a neighboring Qdrant instance, then ran credential sprays against PostgreSQL, SSH and other services. Nothing worked.

A second operation spent more than six hours on a university GPU server. The agents compiled about 80 real usernames by combing faculty directories, lab websites, archived pages and academic papers. They then tried more than 30,000 credential combinations against SSH and, across the campus network, an admin dashboard, a network storage device, a VPN server and the campus portal. No login ever worked.

Both attempts failed. Still, the agents kept adapting for close to a full day without anyone steering them.

The zero-day in a medical imaging app

One target ran an open-source medical imaging AI application, and there the outcome changed. The agents identified the framework, mapped its API endpoints and found an upload endpoint with a filename parameter. They tested it for path traversal, and it was vulnerable. About 11 minutes passed between mapping the API and writing arbitrary files on the server.

In a later session the agents used the same hole to drop a Python reverse shell into sitecustomize.py, a file Python imports automatically at startup. They then called a training API endpoint to launch a fresh Python process. It loaded the payload and opened a reverse shell running as root.

No security advisory, fix, issue or earlier mention in vulnerability databases matched the flaw, so Amazon’s analysts judged with high confidence that it was a zero-day.

A human steps in, and who is behind it

In another session the operator took over. After the zero-day turned up, the actor supplied the agents with a proxy tunneling into a second university network. The server behind that proxy ran the same vulnerable software, which suggests the actor had exploited it to build the tunnel. The agents were then told to target other machines on that network.

Amazon’s analysts attribute the activity to a Chinese-nexus actor with moderate confidence. They point to the Chinese-language prompts, the China Standard Time stamps and the choice of universities that are likely of strategic interest to China.

Fix, notifications and wider use of Cairn

Amazon told the project’s maintainers, who shipped a fix and published a GitHub Security Advisory. The affected universities were not AWS customers, so Amazon had no direct way to reach them. It worked with a government partner to pass the warning along.

Cairn is not limited to one group. Gambit Security has documented a separate, financially motivated actor using it against online retailers, so the framework now appears in both espionage and criminal work.

What defenders should change

An agent that finds and exploits a zero-day in minutes can finish before many security operations centers manually triage the first alert. The agents that failed against Redis ran for 16 hours straight, with no breaks or working-hour limits.

Amazon’s advice is to favor high-confidence detections that trigger automated responses, with people reviewing outcomes instead of every single action. Its own defenses use global honeypots and network traffic analysis to spot actors scanning for weak systems, then help restrict that traffic at the network level at no extra cost to customers.

The indicators listed for the campaign are two IP addresses, 142[.]4.98.145 and 142[.]4.107.34, seen in records from September 8 to 10, 2026.

Source: Amazon Web Services

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.