Eleven minutes to root.
By Ellis Ward | November 21, 2026
Autonomous AI agents found a zero-day vulnerability in a medical imaging platform, exploited it to gain root access on a university server, and did it in roughly 11 minutes. Amazon Threat Intelligence disclosed that campaign on November 20, 2026. The flaw still has no patched release for operators to install.
Agents, a Zero-Day, and a Clock
The disclosure names Cairn, an open-source multi-agent framework, as the tool the attackers used. Amazon researchers found the operation after discovering an unsecured Cairn instance on September 10, 2026. The intrusion itself ran September 8 to 10, according to Amazon Threat Intelligence as reported by VentureBeat.
The target software was MONAI Label, an open-source medical imaging annotation platform developed under Project MONAI. NVIDIA and King's College London co-founded the project. All versions through 0.8.5 carry the vulnerability. Authentication is disabled by default in those versions. Even when operators enable authentication, users with the ANNOTATOR role can still exploit the flaw, according to Amazon Threat Intelligence as reported by VentureBeat.
The campaign targeted university addresses in East Asia outside mainland China. Amazon assessed the operator as China-nexus with medium confidence, citing Simplified Chinese language in prompts, timestamps matching China Standard Time, and the target selection pattern. That assessment does not imply Chinese government involvement.
How the Agents Worked
The operator supplied the target list and a single goal: gain shell access. What followed was largely machine-driven.
Within approximately 11 minutes, the agents mapped MONAI Label's REST API, found an unsanitized filename parameter in a file-upload function, and confirmed they could write arbitrary files through path traversal payloads. That was the zero-day. No prior fix, no open issue, no published advisory, and no database mention had identified it at that point, according to Amazon's high-confidence assessment as reported by VentureBeat.
In a later session, the agents went further. They wrote a reverse shell script into sitecustomize.py, a file Python imports automatically at startup. Then they triggered MONAI Label's training endpoint. That spawned a process that connected back to the attacker with root privileges.
The human operator stepped back in at a later stage. According to BleepingComputer, as cited by CASRAI, threat actors provided a proxy tunnel and directed the agents toward secondary systems across the institutional network. That follow-on step is reported by BleepingComputer and not independently confirmed by Amazon.
A Fix in Code, Not in a Release
The MONAI Label maintainers merged a code fix on October 19, 2026, through pull request 1915. A critical GitHub security advisory, identifier GHSA-qc22-rmgc-vg94, was published on November 18, 2026.
The advisory lists no patched version. The most recent release on the Python Package Index remains 0.8.5, uploaded November 25, 2024, according to VentureBeat. The code fix exists. A compiled release that operators can actually install does not.
For anyone running MONAI Label today, the gap between advisory and installable patch matters. Operators cannot simply update a package. The authentication-disabled default makes exposure worse: a server running MONAI Label out of the box, reachable from outside, sits vulnerable to the same technique the agents used in September.
Speed and Autonomy as the Story
RFTUV has covered cases of AI agents taking unauthorized actions in leasing systems, banking workflows, and workplace software. Those were agents overstepping their intended scope within systems they already had access to. This case sits on a different axis.
Here the agents were directed at a target they had no credentials for and no prior knowledge of. They mapped the attack surface, identified a flaw no security researcher had catalogued, and achieved code execution, all without human direction at the technical level. The 11-minute window is not a rounding error. It reflects how long automated reconnaissance against a REST API takes when nothing slows it down.
Defenders watching for known vulnerability signatures would have found nothing to match. MONAI Label is specialized research software, not a broadly monitored commercial platform. The combination of a niche target, an uncatalogued flaw, and agent-speed discovery narrows the detection window to near zero.
The timeline from exploitation to public advisory also warrants attention. Agents hit the server on September 8. Amazon found the unsecured Cairn instance on September 10 and began its investigation. The fix merged on October 19. The advisory published November 18. As of November 20, 2026, when Amazon went public, no patched release was available on PyPI.
Anyone maintaining a MONAI Label deployment has a narrow set of options right now: take the service offline, restrict network access to trusted hosts, enable authentication, and monitor for path traversal patterns in upload requests. Installing a patched version remains off the table until the maintainers cut a release.

