Update (Nov 21, 2026): MITRE assigned CVE-2026-133968 to the MONAI Label path traversal on November 20, 2026. The official CVE record covers versions through 0.8.5 and rates the issue 9.8 (Critical). As of November 21, no patched release has been published on PyPI or GitHub. The fix remains merged in pull request #1915 and is set to ship in the next MONAI Label release.
A suspected China-linked operator handed a list of university IP addresses to a swarm of AI agents, typed a goal of “obtain shell access” and let the software do the rest. Against one target, the agents found a previously unknown flaw in an open-source medical imaging tool and turned it into a root shell, with no human choosing the next step.
Amazon Threat Intelligence published the findings on November 20, 2026. The agents ran on Cairn, an open-source autonomous penetration-testing framework, and the bug they found sits in MONAI Label, an AI-assisted tool that researchers use to label medical images. Amazon assesses with high confidence that the flaw was a zero-day when it was exploited.
The vulnerability is now tracked as CVE-2026-133968 (GHSA-qc22-rmgc-vg94) and rated Critical. As of November 21, no fixed MONAI Label release was available on PyPI.

An exposed dashboard gave the operation away
Amazon found the campaign on September 10, 2026, when it came across a publicly hosted Cairn instance running without authentication. Because the operator never locked it down, anyone on the internet could read the platform’s history: project configurations, the operator’s prompts and the agents’ output.
The prompts were written in Simplified Chinese, and timestamps in executed scripts were set to China Standard Time. The records showed a campaign aimed at IP addresses tied to universities in East Asia outside mainland China, running for about 48 hours from September 8. Amazon attributes the activity to a “Chinese-nexus” actor with moderate confidence, citing the language, the time zone and the targeting of universities likely of strategic interest to China. It did not tie the operator to a named group.
Amazon said the affected universities were not AWS customers, so it worked through a government partner to notify them.
How Cairn works
Cairn is published on GitHub under the AGPL-3.0 license and describes itself as a general-purpose “state-space search” engine first validated on autonomous penetration testing. Its first public release shipped in May 2026, and the project has since passed 5,000 stars.
The design is simple to operate. A user supplies an origin, such as a target IP address, and a goal, such as getting a shell. Agent workers then read a shared graph of confirmed findings (“facts”) and planned directions (“intents”), decide what to try next, run it and write the results back. A human can inject “hints” at any time. Workers can be backed by Claude Code, Codex or Pi, and the project’s example configuration points them at third-party model APIs.
The project’s documentation says it should only be used where the operator has explicit authorization. Amazon noted that Cairn was built as a research and problem-solving engine for defenders, and that the same ease of use makes it attractive to attackers.
Sixteen hours of failure, then 11 minutes of success
Most of what the agents tried did not work, and Amazon’s write-up is useful precisely because it shows the misses.
The longest operation targeted a university Redis server that allowed access without authentication. Over roughly 16 hours and 148 sequential actions, the agents worked through a catalog of known Redis takeover techniques: abusing replication to load unauthorized code, escaping its scripting sandbox, writing to configuration files and corrupting memory. They also tried Log4Shell against another service on the same host. When those failed, they used a server-side request forgery flaw in a co-located Qdrant instance to map the internal Docker network, then sprayed credentials at PostgreSQL, SSH and other services. Nothing landed.
A second operation ran for more than 6 hours against a university GPU server. The agents built a list of about 80 real usernames from faculty directories, lab websites, archived pages and academic papers, then tested more than 30,000 credential combinations against SSH, an admin dashboard, a network storage device, a VPN server and the campus portal. They never found a working login.
The MONAI Label target went differently. The agents fingerprinted the application, mapped its API and found an upload endpoint that took a filename parameter. They tested it for path traversal and it was vulnerable. Amazon said the agents went from mapping the API to writing arbitrary files on the server in about 11 minutes.
From file write to root shell
In a later session, the agents used the same bug to drop a Python reverse shell into sitecustomize.py, a file that Python imports automatically every time the interpreter starts. They then called a training API endpoint, which spawned a new Python process. That process loaded the payload and opened a reverse shell running as root.
The human operator then stepped in. Amazon said the actor gave the agents a proxy that tunneled into a second university network, landing on a server that ran the same vulnerable MONAI Label software. Amazon infers the operator had already used the zero-day to compromise that machine. The agents were told to use the tunnel to attack other hosts on that network.
Inside GHSA-qc22-rmgc-vg94
The MONAI Label advisory, published November 18, explains the flaw. The PUT /datastore/image endpoint accepts a caller-controlled image query parameter and uses it as the image ID without validation. The datastore code appends a file extension, joins the ID onto the image directory and copies the upload to the resulting path. Because that path is never checked, a value such as ../../../tmp/x writes the file anywhere the MONAI Label process can write. The label upload endpoint, PUT /datastore/label, builds its destination the same way.
Two details make the bug worse in practice:
- Authentication is off by default (
MONAI_LABEL_AUTH_ENABLE=False), so any client that can reach the server can exploit it. - With authentication turned on, the low-privilege
ANNOTATORrole is still enough.
The advisory lists all versions up to and including 0.8.5 as affected and “None” under patched versions. The fix, which confines datastore writes to the expected directories and strips image IDs down to a basename, was merged into the project’s main branch as pull request #1915 on October 19. The maintainers say it will ship in the next MONAI Label release. Amazon first reported the issue privately through the core MONAI repository before it was re-filed against MONAI Label to obtain a CVE.
Not the only Cairn operator
Amazon pointed to earlier research from Gambit Security showing a separate, financially motivated actor using Cairn. In a September 22 report, Gambit said a Chinese-speaking operator chained three open-source AI harnesses (Strix for vulnerability discovery, Cairn for exploitation and Hermes for orchestration) against online retailers. Gambit counted 105 attack projects between September 10 and 15 alone, at least 27 compromised companies and more than 600,000 stolen card records, at a model cost of roughly $25 per target.
Taken together, the two reports show the same free framework being used for espionage and for card theft within weeks of each other. They also add to a fast-growing list of AI-driven intrusions BreachNews has tracked, from the Vapor Panda ARTEX campaign to the AntiFox AI pentesting platform attacks and the 3AM affiliate’s AI agent loop.
What the speed means for defenders
Amazon’s main warning is about tempo. An agent that finds and exploits a zero-day in minutes finishes before many security teams can triage a single alert, and agents that fail keep going around the clock. Amazon recommends prioritizing high-confidence detections that trigger automated response, with people overseeing outcomes rather than reviewing every action.
The failed operations also carry a lesson. Basic hardening held up: the Redis server survived 148 attempts, and the GPU cluster survived 30,000 login guesses. The one success came from an unauthenticated, internet-facing research tool with a bug nobody had found yet.
Steps for MONAI Label users
- Do not expose MONAI Label servers to the internet or other untrusted networks.
- Set
MONAI_LABEL_AUTH_ENABLE=Trueand do not grant theANNOTATORrole to untrusted users. - Run the server as an unprivileged user that can only write to the datastore directory, so a traversal cannot reach
sitecustomize.pyor other code paths. - Look for unexpected
sitecustomize.pyfiles, unknown entries in.ssh/authorized_keysor new cron jobs on hosts that run MONAI Label. - Search logs for
PUT /datastore/imagerequests with../sequences in theimageparameter. - Check network logs for Amazon’s two published indicators, seen between September 8 and 10, 2026:
142[.]4.98.145and142[.]4.107.34. - Upgrade as soon as a MONAI Label release containing pull request #1915 is published.












