AI agents independently find a critical flaw in MONAI Label
Amazon reports observing autonomous AI agents discovering and then exploiting an unknown flaw in MONAI Label, an open-source medical imaging tool. The attack, carried out via Cairn, targeted universities in East Asia outside China.
Amazon claims to have observed in September 2026 a China-linked attack group using autonomous AI agents to discover and then exploit a zero-day vulnerability in MONAI Label, an open-source medical image annotation tool. It took about 11 minutes for the agents to move from API analysis to writing files on the server, before gaining full control of a university machine. The stakes go beyond MONAI Label alone: attackers can now let a machine act almost independently while security teams are still trying to understand what just happened.
Un serveur Cairn exposé a servi de vitrine au mode opératoire
The discovery came from an almost embarrassing detail for the attackers: Amazon Threat Intelligence came across a Cairn instance accessible without authentication. Cairn, an open-source penetration testing framework designed to coordinate multiple AI agents, allows setting a goal in natural language and letting the agents find the method. Here, the operator simply ordered "obtain shell access." The rest was executed by the machine.
According to Amazon, the logs observed showed a campaign lasting about 48 hours, launched between September 8 and 10, 2026, against IP addresses of East Asian universities located outside mainland China. The company attributes the operation with moderate confidence to a China-linked actor, based on the language of the prompts, script timestamps, and target profiles. The story is almost like a bad thriller, except here the server hadn’t even closed the door.
MONAI Label, la cible trouvée sans mode d’emploi
On an initial machine, the agents identified MONAI Label, a medical imaging application used to annotate scans and train AI models. They mapped its API, then spotted a path traversal flaw in a file upload function. This vulnerability allowed writing files to arbitrary locations on the server. Amazon says a switch occurred in just 11 minutes between reconnaissance and exploitation.
The next steps are even more concrete. In a subsequent session, the agents placed a Python reverse shell in sitecustomize.py, a file automatically loaded at interpreter startup, then triggered an application feature to open a root shell. Amazon considers this flaw a zero-day with high confidence, due to the lack of patch, security advisory, or prior mention in vulnerability databases at the time of the events. MONAI Label has since received a patch integrated into its source code, and maintainers published a critical security advisory on November 19 [to be verified] regarding the effective availability of the fixed version.
Les défenseurs courent derrière la machine
The most troubling aspect of this case is not just the speed. It is also the partial autonomy of the campaign. Amazon reports that the agents continued testing other avenues when an attack failed: an attempt against a Redis server lasting 16 hours and 148 actions, a brute force trial with over 30,000 combinations on a university GPU server, and exploration of other services via known vulnerabilities. None of these succeeded, but the persistence alone measures the scale shift.
Amazon also observed that the operator provided a proxy to a second university network where the same vulnerable software was running, suggesting the initial compromise had already served as a pivot. Researchers also cite two IP addresses linked to the campaign: 142.4.98.145 and 142.4.107.34. The key point here is operational: an agent that discovers and exploits a flaw in minutes can outpace a SOC center processing an alert, forcing a strengthening of automated defenses rather than relying solely on manual analysis.
Pourquoi ce cas compte pour les équipes sécurité
Until now, AI-assisted attacks mostly relied on already known flaws or stolen credentials. The MONAI Label case shows something else: an agent can spot an unknown weakness in a niche tool, exploit it, then attempt lateral movement without human intervention at each step. Amazon warned MONAI Label maintainers, who fixed the code, and notified affected organizations via a public partner [to be verified] on this specific notification channel.
The signal is clear for hospitals, universities, and labs exposing servers to the Internet: an open service quickly becomes a target, even when it seems insignificant on paper. Autonomous agents now know how to perform initial reconnaissance, exploitation, and escalation. Not always successfully, of course. But often enough to force security to change pace.
By the numbers
- 11 minutes — from API discovery to file writing, according to Amazon.
- 48 hours — estimated duration of the campaign, from September 8 to 10, 2026.
- 30,000+ combinations — password attempts against a university GPU server.
- 148 actions — attempt to take control of a Redis server over 16 hours.
- 9.8/10 — severity of CVE-2026-133968, according to the MONAI Label security advisory.