HomeTech NewsCybersecurityAWS Says Cairn AI Agents Found and Exploited a Zero-Day

AWS Says Cairn AI Agents Found and Exploited a Zero-Day

Amazon says a Chinese-nexus actor’s Cairn AI agents found a zero-day in an open-source application and used it to take full control of a university server, with little human direction once a goal had been set. In an AWS Security Blog post dated 20 November 2026, Amazon chief information security officer CJ Moses wrote that Amazon Threat Intelligence found a publicly hosted Cairn instance running without authentication on 10 September, exposing the operator’s prompts, project history and output. The campaign, AWS says, ran for about 48 hours beginning 8 September and targeted IP addresses associated with universities in East Asia outside mainland China. The goal given to the agents was to “obtain shell access.”

AWS does not name the application. It describes only “an open-source medical imaging AI application” with a path-traversal flaw in an upload filename parameter. A GitHub Security Advisory for MONAI Label, GHSA-qc22-rmgc-vg94, published on 18 November and originally reported by AWS’s vulnerability disclosure programme, describes a matching critical issue: a caller-controlled image identifier on PUT /datastore/image that can write a file anywhere the process can reach. The Register and BleepingComputer report that this is the application in the campaign. AWS assesses the zero-day finding “with high confidence” and the Chinese-nexus attribution “with moderate confidence.”

How the Cairn AI agents moved from failed attempts to a root shell

AWS describes Cairn as an open-source autonomous penetration-testing framework built as a research tool for defenders. Given a starting point and a plain-language objective, it coordinates multiple agents to plan, execute and adjust without the operator choosing each next step. That ease of use, Moses wrote, also makes it attractive to threat actors. The operator of this instance wrote prompts in Simplified Chinese, and timestamps in executed scripts were set to China Standard Time — two of the points AWS cites for its moderate-confidence nexus assessment, along with the choice of university targets.

Not every operation worked. Against an unauthenticated Redis instance at one university, the agents spent roughly 16 hours and 148 sequential actions on known takeover techniques, then tried a Log4Shell injection against another service on the same system, an SSRF path through a co-located Qdrant instance, and credential sprays against PostgreSQL, SSH and other services. None of those attempts succeeded. In a separate operation lasting more than six hours, they compiled about 80 real usernames from public university material and tested more than 30,000 credential combinations against a GPU server and other campus systems. They never found a working login. “Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction,” Moses wrote.

The medical-imaging application was where the Cairn AI agents succeeded. AWS says the agents mapped the application’s API, found an upload endpoint with a filename parameter, and tested it for path traversal. “In about 11 minutes, the agents went from mapping the application’s API to arbitrarily writing files on the server.” In a later session they used the same flaw to write a Python reverse shell to sitecustomize.py, a file Python imports at start-up, then triggered a training API endpoint so a new process imported the payload and opened a reverse shell running as root. At the time of discovery, AWS says, the flaw had no assigned advisory and was unpatched. After the agents found it, the human operator provided a proxy into a second university network running the same vulnerable software and directed the agents to target other machines there.

The affected organisations were not AWS customers, and AWS says it had no direct path to notify them. It worked with a government partner to pass on the warning, and it notified the project’s maintainers. AWS also published two internet addresses seen in records dated 8 to 10 September 2026.

This is not the first time Cairn AI agents have appeared in offensive operations. TSN has already reported Reuters’ account of Cairn agents used against more than 100 company websites, a case AWS itself cites via Gambit Security as a financially motivated contrast to this university campaign. It also sits close to GTIG’s report of fully automated vulnerability exploitation, and to the FBI FLASH on ARTEX, another AI agent framework that US officials have described in defensive terms.

A pull-request fix is not a shipped MONAI Label release

Amazon’s post says the maintainers “released a fix and a GitHub Security Advisory.” The advisory itself is more precise, and that precision matters for anyone still running the software. GHSA-qc22-rmgc-vg94 lists monailabel 0.8.5 and earlier as affected, gives no patched package version, and has no CVE yet. It says the fix is in pull request 1915 and “will be included in the next MONAI Label release.” Until that release exists, the practical defence against Cairn AI agents — or any other client that can reach the server — is the workaround list in the same advisory: turn on authentication with MONAI_LABEL_AUTH_ENABLE=True and keep the ANNOTATOR role away from untrusted users; keep the server off untrusted networks; and run it as an unprivileged user whose write access is limited to the datastore directory. Authentication, the advisory says, is off by default, so any network client that can reach a default install can try the write.

The 11-minute success of the Cairn AI agents should not be read in isolation from the failed day that preceded it. The Redis and GPU-server operations did not yield a shell, but they did keep going for many hours, shifting technique when a path failed. The lesson AWS draws for defenders is about tempo: an agent that finds and exploits a flaw in minutes can finish faster than a security team can triage an alert by hand, and a failed run can still occupy a target for most of a day. High-confidence, automated response, Moses argues, has to carry more of that load.

Two caveats should stay attached to the headline. The Chinese-nexus judgement is moderate, not high, and rests on language, timestamps and target choice rather than on a named group. And the application name used in this report is MONAI Label because of the advisory and because The Register and BleepingComputer link that advisory to the campaign, not because Amazon’s blog names the product. Operators should treat the workarounds as the current advice, and watch for the next MONAI Label release rather than assume a patched package is already out.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.