Amazon: a China-linked actor's Cairn agents found a MONAI Label zero-day on their own and went from API map to file write in 11 minutes
Amazon Threat Intelligence says autonomous agents told only to 'obtain shell access' found an unknown path traversal in MONAI Label and got a root shell on a university server. The fix is merged, but no patched release has shipped.
On November 20, 2026, Amazon published an AWS Security Blog post by CJ Moses, CISO of Amazon Integrated Security. It describes a campaign in which an operator handed a list of IP addresses to Cairn, an open-source multi-agent penetration-testing framework, with variations of one instruction: “obtain shell access.” The agents planned and adjusted their attacks with no step-by-step direction. Against a server running MONAI Label, an open-source medical imaging annotation tool, they found a path traversal bug that had no advisory, issue or fix anywhere. According to Amazon, it took them “about 11 minutes” to go from mapping the application’s API to writing arbitrary files on the server.
Earlier agentic intrusions mostly chained known bugs and stolen credentials. This report has agents finding and weaponizing a genuinely unknown flaw in a real operation. That changes what defenders have to plan for. The fix also matters for anyone running MONAI Label: it was merged on October 19, but as of today no release that contains it has been published.
What Amazon saw
Amazon could reconstruct the campaign because the operator left their Cairn server open to the internet with no authentication. On September 10, Amazon Threat Intelligence found it. Anyone could read the project configurations, the operator’s prompts and every piece of agent output. The prompts were written in Simplified Chinese and the timestamps in executed scripts were set to China Standard Time. The campaign ran for about 48 hours from September 8. It targeted IP ranges belonging to universities in East Asia outside mainland China. Amazon attributes it to a Chinese-nexus actor with moderate confidence. The basis is the language, the time zone and target selection that fits Chinese strategic interest. No named group is identified.
The log shows three operations that ended very differently:
- A hardened Redis server: about 16 hours, 148 sequential actions, no success. The Redis instance accepted unauthenticated connections, and the agents worked through the standard takeover playbook. They tried replication abuse to load a malicious module, a Lua sandbox escape, configuration file writes and memory corruption, plus Log4Shell injection against another service on the host. When all of that failed, they used a server-side request forgery bug in a co-located Qdrant vector database to map the internal Docker network. Then they sprayed credentials at PostgreSQL, SSH and other services. Nothing worked.
- A university GPU cluster: more than 6 hours, no success. The agents built a list of about 80 real usernames from faculty directories, lab pages, archived web pages and academic papers. They then tried more than 30,000 credential combinations against SSH, an admin dashboard, a NAS, a VPN and the campus portal. They never found a working login.
- MONAI Label: a zero-day and a root shell. The agents identified the framework, enumerated its API, and noticed an upload endpoint that took a filename parameter. They tested that parameter for traversal, and it was vulnerable.
Amazon’s own summary of the failures is the important part. Neither failed operation produced access, but the agents “kept adapting for nearly a full day without human direction.” A human operator would have taken breaks or given up. The agents did neither.
How the MONAI Label exploit worked
The bug is documented in GitHub advisory GHSA-qc22-rmgc-vg94, published November 18 and rated critical, with no CVE assigned yet. In MONAI Label 0.8.5 and earlier, PUT /datastore/image takes a caller-controlled image query parameter and uses it as the image ID without validation. LocalDatastore.add_image() appends a file extension, joins the ID onto the datastore directory and copies the upload there with shutil.copy(). It never checks that the result stays inside the datastore. As a result, a value like ../../../tmp/x writes the file anywhere the process can write. PUT /datastore/label builds its path the same way. Two defaults make this worse. Authentication is off by default (MONAI_LABEL_AUTH_ENABLE=False), so anyone who can reach the port can exploit it. With authentication on, the low-privilege ANNOTATOR role is still enough.
Amazon describes how the agents turned the file write into code execution in a later session. They wrote a Python reverse shell into sitecustomize.py, a file the Python interpreter imports automatically at startup. They did not wait for a restart. They called MONAI Label’s training API, which spawns a new Python process. That process imported the payload and connected back as root, because the server ran as root on GPU hardware.
The operator then stepped in for the only time the log records. They gave the agents a proxy that tunneled into a second university’s network. The machine at the other end of that tunnel was running the same vulnerable MONAI Label. Amazon infers the actor had already used the zero-day to compromise it, and then sent the agents to attack other machines on that network. Amazon notified the maintainers. Because the victims were not AWS customers, it worked with a government partner to notify them.
The patch gap
Amazon’s post says the maintainers “released a fix and a GitHub Security Advisory.” That is true at the code level, but not for anyone installing from a package index. The fix is pull request #1915, “fix: confine datastore write paths,” merged into main on October 19. It resolves both destinations and rejects anything outside the datastore directories. It also reduces the image ID to its basename and rejects empty or dot-prefixed IDs with HTTP 400. However, the advisory lists the patched version as blank and says the fix “will be included in the next MONAI Label release.” PyPI still shows 0.8.5 as the latest version, a release from November 2024. A Cloud Security Alliance research note published today points out that the fix has now been public in the repository for more than a month without an installable release. Over that time, anyone diffing the repo could see exactly what was fixed.
Until a release ships, the advisory’s workarounds are the practical defense:
- Set
MONAI_LABEL_AUTH_ENABLE=Trueand keep theANNOTATORrole away from untrusted users. - Take MONAI Label servers off untrusted networks.
- Run the server as an unprivileged user that can write only to the datastore directory. That alone would have stopped this chain from reaching root.
To hunt for past exploitation, check logs for PUT /datastore/image or /datastore/label requests containing ../, encoded traversal or absolute paths. Inspect sitecustomize.py, usercustomize.py and .pth files in the server’s Python environments, and check authorized_keys and cron. Amazon lists two IP addresses seen from September 8 to 10: 142[.]4.98.145 and 142[.]4.107.34. Teams that cannot wait can build from a commit that includes PR #1915. Note that the PR also changes CI and dependency pins, so test it in staging first.
Cairn keeps showing up
Cairn is not new to this blog. In October’s $25-per-breach campaign, it was the exploitation stage of a three-framework stack used by a single operator. Cloudflare’s November report on 50+ AI pentest framework families singled it out as a dispatcher that drives existing coding-agent harnesses rather than running its own agent loop. Amazon credits Gambit Security as one of the first to document a threat group using it: a financially motivated actor targeting online retailers. Amazon says the two cases together show the framework “is being adopted in both espionage and criminal operations.” Cairn was built as a research and defense tool, and Amazon’s post makes the uncomfortable point that its ease of use is exactly what makes it useful to attackers. The input is a starting point and a goal in plain language.
The report leaves several things open. Amazon does not say which models the operator’s Cairn instance was using. It does not say how many sessions or how much compute the campaign consumed. It also cannot know whether the operator ran other instances that stayed hidden. A single exposed panel is all anyone saw.
What it means for defenders
The tempo is the headline. An agent that goes from API discovery to arbitrary file write in 11 minutes finishes before most security operations centers would have triaged the first alert. Moses’s recommendation follows from that: rely on high-confidence detections that trigger automated responses, with humans overseeing outcomes instead of reviewing each action. The Redis and GPU-cluster logs show the other half of the problem. Hardening and strong credentials still worked here. But the attacker’s cost of trying for 16 hours is now compute, not a person’s time. A low per-target success rate is still worth running across a whole IP range.
The target choice is a warning too. Research AI tooling, such as annotation servers, notebook hosts, model servers and vector databases, is often set up by small lab teams. They tend to use default settings, run it as root on GPU machines and keep no asset inventory. MONAI Label was the soft spot in this campaign because it shipped with authentication off. Any organization running similar tools should find them, check whether they are reachable from the internet and look at the user they run as. A machine is already better at finding these servers than most of the people who run them.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
ADK Adds Native Live Evaluation for Voice Agents
Learn how to evaluate real-time voice agents in Google ADK with audio simulation, trajectory rubrics, CLI automation, and ADK Web playback.
One Operator, Three Agent Frameworks: Inside the $25-per-Breach AI Hacking Campaign
Reuters reported October 19 that a single hacker used autonomous AI agents against 1,200+ sites including Deere, SharkNinja, and Cummins, and a Cloud Security Alliance note reconstructs the full stack: three open-source agent frameworks running 27+ breaches at an average of $25.46 per company.
Google found exploit_agent: a framework that diffs patches, writes the exploit and fires it with no human involved
GTIG's Q4 2026 AI Threat Tracker reports the first fully automated patch-to-exploit pipeline seen on an attacker's server, plus LLM Mixers reselling stolen model access and a Claude Code settings.json backdoor. What the evidence supports and what to change.
Cloudflare: 50+ AI pentest frameworks scanned its customers in three days, and traffic is up 30x since October 1
Cloudflare's November 13 report says taking ARTEX offline produced a hydra effect: hundreds of forks, 50+ framework families and a 30x rise in agentic attack traffic. Here is what the data shows, what it leaves out, and what a stolen CI/CD token looked like in one incident.
ARTEX in the wild: Chainstack's breach and the AntiFox fork show what agentic attack frameworks actually do
Chainstack says an attacker using the ARTEX agent framework went from a default-password Grafana to production secrets in five days, and Arctic Wolf ties an ARTEX fork called AntiFox to AWS secret theft. Here is how the intrusions worked and what defenders should look for.