Amazon Threat Intelligence revealed on Friday, November 20, 2026, that autonomous AI agents found an unknown security flaw — a zero-day — on their own in MONAI Label, an open-source server research teams use to label medical scans and train AI models, and used it to gain root access on a university server.
The case came to light because the operator left their own Cairn control panel exposed on the internet without authentication. On September 10, 2026, Amazon found the Cairn instance — an open-source framework that coordinates AI agents as autonomous penetration testers — with its settings, prompts and run history visible to anyone. The campaign lasted about 48 hours, starting September 8, against IP addresses associated with universities in East Asia, outside mainland China.
How the agents found the flaw
The agents were given a single instruction in plain language — "obtain shell access" — plus a list of targets. From there, they planned, executed and adjusted every step on their own. Against the server running MONAI Label, they mapped the application's API, found an upload endpoint with an unvalidated filename parameter and tested it for path traversal, a technique that uses sequences such as "../" to escape the intended folder. In about 11 minutes, they went from mapping the API to writing arbitrary files on the server.
Root access came in a later session: the agents wrote a reverse-shell payload into sitecustomize.py, a file Python imports automatically at startup, and triggered a MONAI Label training endpoint. The new Python process, started as root, loaded the file and opened a connection back to the attacker.
Not everything worked for the agents. The same campaign included an operation of about 16 hours and 148 actions against an unauthenticated Redis server, and another of more than 6 hours against a university GPU server, testing over 30,000 credential combinations without finding a valid login. Both failed — but they show agents persisting for almost a full day without human direction.
Amazon assesses with moderate confidence that the actor is China-linked, based on prompts in Simplified Chinese, timestamps in China Standard Time and targets likely of strategic interest to China. No group was named and no government link was established. The assessment that the flaw was a zero-day carries high confidence: there was no fix, advisory or prior mention in any vulnerability database.
The flaw, and the fix that has not shipped yet
The flaw is a path traversal in the MONAI Label datastore upload endpoints and affects versions 0.8.5 and earlier in the default setup, which runs without authentication. It received the identifier CVE-2026-133968, with a CVSS score of 9.8, rated critical. The maintainers confirmed the flaw, merged the fix into the source code on October 19, 2026, and published a security advisory this week — but the newest version available to install is still 0.8.5, from November 2024. In other words, the fix exists in code, but there is no patched release to install yet.
What to do until a patched release ships
Until a release ships with the fix, the recommendation is to block untrusted network access to MONAI Label servers, turn on authentication (off by default), run the server under a limited account instead of root, and check for unexpected Python startup files and unusual training jobs.
There are important limits to the account: it is Amazon's own reconstruction and has not been independently confirmed. There is no sign that patient scans, records or clinical systems were accessed. Root access was confirmed on one server; on a second university network, reached through a proxy tunnel supplied by the operator, Amazon believes the same method was used, but the compromise was not confirmed.
The case is not isolated: Gambit Security had already documented another actor using Cairn against online retailers. The framework, built as a defensive tool, now shows up in espionage and financially motivated operations — and the agents' speed, finding and exploiting a flaw in minutes, compresses defenders' response window.
It is not the first time an AI has found flaws on its own: in September, OpenAI rated its Astra AI as "Critical" in cybersecurity after it discovered zero-day flaws alone. Read: OpenAI rates its Astra AI as "Critical" in cybersecurity after it discovered zero-day flaws on its own.
Image credit: illustration by The Hacker News about the MONAI Label attack.
Sources: AWS Security Blog — Chinese-nexus actor's Cairn AI agents found a zero-day on their own; The Hacker News — Attackers' AI Agents Found a Zero-Day in Medical-Imaging Software and Gained Root Access; VentureBeat — Amazon says attackers' AI agents found and exploited a zero-day in 11 minutes, but no patched release is available; Security Affairs — Cairn AI Framework Used in Autonomous Zero-Day Exploitation.
Transparency: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and sources on the internet. Always consult the original sources listed above.
By Muse