Just in
TechUbisoft Motherbrain AI Unit Launches as H1 Net Bookings Fall 16.5%AIChatGPT Space Rolls Out to Plus Users, Replacing the Library TabTechAnthropic IPO Closes With $119.5 Billion in Combined Gross ProceedsAIRiemann Hypothesis Proof Certified by Lean and Mathlib CommunitiesTechCairn AI Agents Found a Zero-Day in MONAI Label, Amazon SaysAITasks in Gemini Now Calls Businesses for You, With Duplex Inside
Tech

Cairn AI Agents Found a Zero-Day in MONAI Label, Amazon Says

Amazon says Cairn AI agents run by a suspected China-linked operator found an unknown flaw in the MONAI Label medical imaging tool and gained root access on a university server. Version 0.8.6 fixes the bug.

USA Publishing Group5 min read
Share
Lines of source code on a dark computer screen, the kind of software Cairn AI agents probed to find a zero-day flaw
Photo: Chris Ried / Unsplash

Key takeaways

  • Amazon Threat Intelligence said Cairn AI agents found a previously unknown path traversal flaw in MONAI Label and used it to get a root shell on a university server.
  • The MONAI Label advisory rates the bug, tracked as CVE-2026-133968, critical for version 0.8.5 and earlier, and version 0.8.6 fixes it along with a second upload flaw.
  • Amazon attributed the September campaign against East Asian universities to a China-linked operator with moderate confidence and urged automated responses to high-confidence detections.
In this story 3 sections
  1. How the Cairn campaign unfolded
  2. The MONAI Label flaw and the fix
  3. Why Amazon says defenders need automation

An AI hacking framework called Cairn, left running on the open internet with no login, gave Amazon's security team a rare look at what autonomous AI agents can do when someone points them at real targets. According to a post on the AWS Security Blog published Friday, Nov. 20, Cairn AI agents working for a suspected China-linked operator discovered a previously unknown flaw in an open source medical imaging tool and used it to get a root shell on a university server.

Amazon (NASDAQ: AMZN) did not name the software in its write-up, but the post links to a security advisory from the MONAI Label project, which describes itself as an open source tool for labeling medical images and training AI models on them. A patched version, MONAI Label 0.8.6, is dated Nov. 22 on the project's GitHub release page.

The post is signed by CJ Moses, chief information security officer of Amazon Integrated Security. Amazon said it holds "moderate confidence" in the China link, citing prompts written in Simplified Chinese, timestamps set to China Standard Time and a target list of universities. It did not name a specific group.

How the Cairn campaign unfolded

Cairn is an open source penetration testing framework. A user gives it a starting point and a goal in plain language, and it directs several AI agents to work toward that goal. Amazon's account, put in order:

  1. Sept. 8, 2026: The operator fed Cairn the IP addresses of universities in East Asia outside mainland China, along with variations of one instruction: get shell access. The campaign ran for about 48 hours.
  2. Redis attempt: Against one university's Redis database, which had no authentication, the agents worked for about 16 hours and took 148 actions in sequence. They tried replication tricks, a Lua sandbox escape, config file writes, memory corruption and a Log4Shell injection on another service, then used a flaw in a Qdrant instance on the same host to map the internal Docker network and tried passwords against PostgreSQL and SSH. Amazon said none of it worked.
  3. GPU server attempt: Over more than six hours, the agents pulled about 80 real usernames from faculty pages, lab sites and research papers, then tried more than 30,000 username and password pairs against SSH, a VPN, a campus portal and other systems. No login succeeded.
  4. The zero-day: On a server running the medical imaging tool, the agents mapped its API, found an upload endpoint that accepted a file name without checking it, and abused it to write files anywhere on the machine. Amazon said about 11 minutes passed between mapping the API and the first arbitrary file write.
  5. Root shell: In a later session, the agents planted a Python reverse shell in a file that Python loads at startup, then called the tool's training endpoint to launch a new Python process. That gave them a shell running as root.
  6. Sept. 10, 2026: Amazon Threat Intelligence found the exposed Cairn server, with its project settings, the operator's prompts and the AI output all visible.

After the agents got in, the human operator supplied a proxy into a second university network, where a server ran the same vulnerable software. Amazon said it believes the operator had already compromised that machine with the same flaw, and the operator then told the agents to go after other systems on that network.

By our count, the credential attack worked out to roughly 375 password guesses for each of the 80 usernames (30,000 divided by 80), and 69 days passed between Amazon spotting the Cairn server on Sept. 10 and the project's advisory going public on Nov. 18.

The MONAI Label flaw and the fix

The MONAI Label advisory, GHSA-qc22-rmgc-vg94, rates the bug critical. It affects version 0.8.5 and earlier. The datastore upload routes, PUT /datastore/image and PUT /datastore/label, used a caller supplied ID as part of a file path, so an ID containing "../" could land a file outside the intended folder. The advisory says that could let an attacker overwrite program code, SSH keys or scheduled jobs, or knock the server offline. Authentication is off by default, and even with it on, the basic annotator role is enough to exploit the bug. The issue was first reported privately by AWS's vulnerability disclosure team.

The MONAI Label 0.8.6 release notes list the bug as CVE-2026-133968 and also fix a second path traversal issue in session uploads that unpacked tar archives without checking file names. The project says anyone who ran 0.8.5 or older on a network reachable by untrusted users should check for unexpected file changes. Its other guidance:

  • Upgrade with pip install -U monailabel==0.8.6, or pull the projectmonai/monailabel:0.8.6 Docker image.
  • Keep the server off untrusted networks.
  • Turn on authentication and limit who gets the annotator role.
  • Run the service as an unprivileged user that can only write to its datastore.

Amazon said it worked with a government partner to warn affected organizations that are not AWS customers. It published two IP addresses tied to the activity between Sept. 8 and Sept. 10: 142.4.98.145 and 142.4.107.34.

Why Amazon says defenders need automation

Amazon assesses "with high confidence that it was a zero-day," because no fix, advisory or public record of the bug existed when the agents found it. Its broader point is about speed: an agent that can turn API mapping into a working exploit in minutes, and other agents that keep trying for 16 hours without a break, can outrun a security team triaging alerts by hand. The company recommends automated responses to high-confidence detections, with people overseeing the results instead of approving each step.

The report adds to a run of findings about AI agents in real attacks, including Amazon's earlier report on ARTEX AI agents used in key theft and deepfake attacks and Google's AI Threat Tracker on exploits run with no human in the loop. Amazon also noted that Gambit Security has documented a separate, financially motivated group using Cairn against online retailers. More security coverage is on our tech page.