News AI Security
Amazon: Attacker-Run Cairn AI Agents Found and Exploited a MONAI Label Zero-Day
Amazon Threat Intelligence says an actor pointed the open-source Cairn agent framework at a list of IP addresses with the goal "obtain shell access", and the agents found and exploited an unknown path traversal in MONAI Label. Two other targets in the same campaign held.

Key takeaways
- Amazon says attacker-run Cairn agents went from API mapping to arbitrary file write on one target in about 11 minutes.
- The flaw is a path traversal in MONAI Label upload endpoints, rated Critical, with no known CVE.
- The maintainer advisory lists no patched release yet, the fix ships in the next release.
- Authentication is off by default, which made the default configuration the real exposure.
- Two other targets absorbed hours of agent effort and held.
Amazon Threat Intelligence has documented a campaign in which autonomous AI agents, not a human operator, found and exploited a previously unknown vulnerability in a real target.
What Amazon reported #
According to the AWS Security Blog, Amazon identified a publicly hosted instance of Cairn running without authentication, which exposed project configurations, the actor's prompts and the AI output. Amazon describes Cairn as an open-source autonomous AI penetration-testing framework that was built as a research and problem-solving engine for defenders.
The campaign ran for about 48 hours in September against IP addresses belonging to universities in East Asia outside mainland China. The operator gave the agents a list of targets and one goal: "obtain shell access".
On one target, Amazon says, "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server." That figure is Amazon's, and it covers API mapping to arbitrary file write on that single target, not the whole campaign. The reverse shell was written in a later session.
The bug #
Amazon's post calls the software an open-source medical imaging AI application and links to the Project MONAI advisory GHSA-qc22-rmgc-vg94 for MONAI Label. The advisory rates the issue Critical and classifies it as CWE-22. The datastore image upload endpoint uses a caller-controlled filename parameter without validation, and the label upload endpoint builds its destination the same way. The result is arbitrary file write.
The advisory states there is no known CVE. It lists no patched release yet: the fix has been merged and will ship in the next MONAI Label release.
Authentication is disabled by default, so in the default configuration any network client that can reach the server can exploit the flaw. With authentication on, the annotator role is enough.
What did not work #
The same campaign shows the limits. Amazon says the agents spent about 16 hours and 148 sequential actions on a Redis target, and more than 6 hours and more than 30,000 credential combinations on a university GPU server. Both attempts failed.
Attribution #
Amazon assesses with moderate confidence that the operator is a Chinese-nexus actor, citing Simplified Chinese prompts, China Standard Time timestamps and the choice of targets. It names no group. It assesses with high confidence that the flaw was a zero-day, because it found no corresponding fix, issue, advisory or database entry. Amazon says the affected organisations were not AWS customers and were notified through a government partner.
What to do #
- Enable authentication on MONAI Label, which is the advisory's first workaround.
- Do not expose the server to untrusted networks.
- Run it as an unprivileged user with write access limited to the datastore directory.
- Treat research and AI tooling as production software in code review. An unvalidated filename in an upload handler is a finding standard appsec review catches.
Amazon's own recommendation is to "prioritize high-confidence detections that trigger automated response, with humans overseeing outcomes".
Sources
- Chinese-nexus actor's Cairn AI agents found a zero-day on their own, AWS Security Blog
- Path traversal in datastore upload endpoints allows arbitrary file write, Project MONAI (GitHub Security Advisory)
- Amazon: China-Linked Hackers' AI Agents Exploited Medical Software Zero-Day, SecurityWeek
Questions this raises
- Is there a CVE for the MONAI Label flaw?
- No. The maintainer advisory GHSA-qc22-rmgc-vg94 states there is no known CVE.
- Is a patched MONAI Label release available?
- The advisory lists no patched release. It says the fix has been merged and will be included in the next MONAI Label release. Until then, enable authentication, keep the server off untrusted networks and run it as an unprivileged user.
- Who was behind the campaign?
- Amazon assesses with moderate confidence that it was a Chinese-nexus actor and does not name a group. That is Amazon's assessment, not an established fact.
- Did the agents succeed everywhere?
- No. Amazon reports two failed attempts in the same campaign, one against a Redis target and one against a university GPU server.