Cairn found a zero-day and took control of a university server

Cairn found a zero-day and took control of a university server

Artificial intelligence agents found a zero-day vulnerability in MONAI Label and used it to execute a shell as root on a university server. Amazon Threat Intelligence says the system went from analyzing the application’s API to writing arbitrary files in about 11 minutes; afterward, the attackers reused the flaw to enter another university network.

Amazon detected the operation on September 10, 2026, after finding a Cairn instance exposed to the Internet without authentication. Its public history included instructions, configurations, and agent results. The campaign had begun on September 8 and lasted about 48 hours, with IP addresses linked to universities in East Asia outside mainland China.

An easy target for autonomous agents

Cairn is an open-source penetration-testing framework, created as a research tool and to help strengthen defenses. The operator provided IP addresses and a goal expressed in plain language: gain access to a shell. From there, the agents organized and adjusted their attempts without the person choosing each step.

“Based on those inputs, the AI agents planned, executed, and adjusted their approach as they learned, without the operator choosing each next step”

— CJ Moses, CISO of Amazon Integrated Security

The attribution is not definitive. Amazon assesses with moderate confidence that this was an actor linked to China, based on instructions written in Simplified Chinese, timestamps set to China Standard Time, and the selection of universities that, according to the company, could be of strategic interest to the country. Amazon and AWS did not identify the institutions or the specific country where they were located.

Two failed attempts before the discovery

The campaign also shows that autonomy did not guarantee success. In one operation, the agents spent around 16 hours and 148 actions trying to compromise a university’s Redis server. They tested techniques against the service and other network components, but failed to get in.

Among the actions described by AWS were abusing Redis replication to load unauthorized code, escaping the script sandbox, modifying configuration files, and corrupting memory. When those attempts failed, the agents tried a Log4Shell injection against another service and a server-side request forgery (SSRF) vulnerability in a nearby Qdrant instance. They also tested credentials against PostgreSQL, SSH, and other services, without success.

Another operation targeted a university GPU server for more than six hours. The agents gathered about 80 real usernames from public directories, lab websites, archived pages, and academic papers, and tested more than 30,000 credential combinations. Targets included remote SSH access and, on other systems in the network, an administration panel, network-attached storage, a VPN, and the campus portal. They found no valid credentials either.

A file path opened the server

The attack that succeeded targeted MONAI Label, an open-source medical imaging application that helps researchers label images using artificial intelligence tools. The agents identified the application framework, mapped its API endpoints, and found an upload function with a parameter for the filename. When they tested it, they discovered that it accepted a manipulated path.

The problem lay in the validation of write paths. According to the security advisory published by the MONAI project, the image upload endpoint accepted an identifier controlled by the requester and used it without checking that the destination remained within the data directory. A value containing directory traversal sequences could cause the server to write the file to arbitrary locations accessible to the process.

In a later session, the agents wrote a reverse shell payload to the sitecustomize.py file, which Python automatically imports at startup. They then triggered a training API endpoint to start another Python process; that process loaded the malicious payload and opened a callback connection with root privileges. AWS assessed with high confidence that it was a zero-day: when it was discovered, there was no patch, advisory, or prior mention of the flaw in the project’s vulnerability databases.

The GitHub advisory explains that with authentication disabled—the default configuration indicated there—any network client able to reach the server could attempt to exploit the flaw. If authentication was enabled, the ANNOTATOR role was sufficient. The fix makes the program resolve destinations and reject paths outside the permitted directories; it also normalizes the image identifier and returns an HTTP 400 error for empty values or values beginning with a dot.

The same flaw reached another university

After locating the zero-day, the operator intervened and provided the agents with a proxy that entered a second university network. It had a server running the same vulnerable software, and the actor instructed them to look for other machines inside that network. AWS infers that the operator had previously used the flaw to compromise that server and establish the tunnel.

Amazon reported the vulnerability to the MONAI Label maintainers. The project fixed the flaw in the change identified as PR #1915 and published a security advisory on GitHub. BleepingComputer reported that the security update was released on November 18. AWS said the affected institutions were not AWS customers and that it therefore worked with a government partner to notify them about the activity.

A tool used in different operations

The case adds to other documented uses of Cairn. Gambit Security described a credit card data theft operation targeting online retailers in September. The Register reported that the activity may have affected 27 companies to varying degrees, as part of a campaign that also used two other AI agent frameworks. Amazon presents both episodes as evidence that Cairn appears in both espionage operations and financially motivated activity.

Google had also reported the use of Cairn in an intrusion against a U.S. aerospace organization. Separately, The Record reviewed warnings about Artex, another AI-based offensive tool, and cited an FBI advisory noting its ability to automate several stages of an intrusion. These are different tools: the university incident described by Amazon involved Cairn.

Amazon warned that agents can reduce the time available to respond: in this case, discovery and exploitation occurred within minutes, while other attempts remained active for hours without breaks. AWS recommends prioritizing high-confidence detections that trigger automated responses, with human oversight of the results, rather than manually reviewing every action. The company also published two IP addresses associated with the campaign, observed between September 8 and 10, 2026: 142[.]4.98.145 and 142[.]4.107.34.

Sources: BleepingComputer, The Record, theregister, Amazon Web Services, GitHub, GitHub