On November 20, 2026, Amazon Threat Intelligence published evidence that autonomous Cairn agents discovered and exploited an unpatched path traversal in an open-source medical-imaging stack, gained root on a university server, and only needed a one-line goal: obtain shell access. GitHub advisory GHSA-qc22-rmgc-vg94 identifies the application as MONAI Label. The campaign ran September 8–10, 2026 against East Asian universities outside mainland China; researchers recovered it because the operator left a public, unauthenticated Cairn console on the internet. If you run coding agents that can install offensive harnesses, touch medical-imaging lab stacks, or leave agent UIs on a laptop exposed through a tunnel, this is the speed and persistence model to plan against—not a hypothetical red-team slide.
Amazon CISO CJ Moses''s post is the primary source. Beam previously covered Unit 42''s Cairn CL-STA-1238 field analysis (classic vulns at scale) and exposed Cairn consoles on the internet. This article is the zero-day university chain: failed Redis and credential operations, 11-minute MONAI exploitation, sitecustomize.py persistence, and GHSA-qc22-rmgc-vg94.
TL;DR
| Question | Answer |
|---|---|
| Am I affected? | Yes if you run MONAI Label 0.8.5 and earlier reachable from untrusted networks, especially with default auth off (MONAI_LABEL_AUTH_ENABLE=False) |
| Is there a fix? | Merge #1915 / wait for next release; until then use auth on, network isolation, unprivileged service account per the advisory |
| Did agents find the bug alone? | Amazon assesses high confidence it was a zero-day at exploitation time; maintainers published GHSA-qc22-rmgc-vg94 on November 18, 2026 |
| Does this hit my laptop agent? | Indirectly—not as Cairn malware on your machine, but as the same harness class some developers clone for tests, and as Python services agents deploy to GPU servers |
| What should I hunt? | Unauthenticated Cairn/ARTEX UIs, MONAI Label datastore uploads, IOCs 142.4.98.145 and 142.4.107.34 (Sept 8–10, 2026) |
How Amazon discovered the campaign
On September 10, 2026, Amazon Threat Intelligence found a server hosting Cairn—an open-source autonomous AI penetration-testing framework—without authentication. Anyone could read project configurations, operator prompts, and model output. Prompts were in Simplified Chinese; script timestamps used China Standard Time. The visible history covered roughly 48 hours from September 8 against university IP addresses in East Asia outside mainland China.
Cairn is built for defenders: give a starting point and plain-language objective, and it coordinates multiple AI agents to find a path. Amazon notes that same simplicity attracts threat actors. In this campaign the actor supplied targets plus obtain shell access variants; agents planned, executed, and adapted without the operator picking each micro-step—until a human stepped in after the zero-day to provide a proxy into a second university network.
Discovery through an exposed operator console is not unique. Beam''s GitGuardian catalog of agentic pentest frameworks and the FBI FLASH on ARTEX misuse both stress unauthenticated control planes as a recurring leak vector—whether the attacker forgets a firewall rule or a developer tests from a --host 0.0.0.0 bind.
Attack chain step by step
Phase 1: Long failures on hardened services
Amazon highlights two failed but instructive threads:
Redis at one university (≈16 hours, 148 actions). The instance allowed access without authentication. Agents chained replication abuse, Lua sandbox escapes, config writes, and memory corruption tricks, then tried Log4Shell on a co-located service. When that failed, they used SSRF through a Qdrant instance to map Docker internal networks and sprayed credentials against PostgreSQL, SSH, and other services. Nothing succeeded.
GPU server (6+ hours). Agents harvested ~80 real usernames from public faculty pages, papers, and archives, then tested 30,000+ credential pairs against SSH, an admin dashboard, NAS, VPN, and a campus portal. No valid login.
Amazon''s framing: "The agents kept adapting for nearly a full day without human direction." That persistence matters for SOC triage—your alert queue is not paced to human lunch breaks anymore.
Phase 2: MONAI Label mapping to arbitrary write (≈11 minutes)
Against a host running MONAI Label—the open-source medical imaging annotation stack—agents:
- Identified the framework
- Mapped API endpoints
- Found an upload endpoint with a caller-controlled
filenameparameter (the GHSA documents it as theimagequery parameter onPUT /datastore/image) - Tested path traversal on that parameter—it worked
- Moved from API map to arbitrary file write in about 11 minutes
Phase 3: Root shell via Python startup hook
In a later session, agents wrote a Python reverse-shell payload to sitecustomize.py, which Python imports at interpreter startup. They triggered a training API endpoint to spawn a new Python process; that process imported the payload and opened a reverse shell as root.
Phase 4: Human operator widens scope
After the zero-day, the human operator supplied a proxy tunnel into another university network where the same vulnerable software ran. Amazon infers the operator had already compromised the tunnel endpoint with the same flaw and aimed agents at additional machines inside that network.
GHSA-qc22-rmgc-vg94: who is affected and how to fix
GitHub advisory GHSA-qc22-rmgc-vg94 (published November 18, 2026, critical severity) tracks the flaw in the MONAILabel repository—the code lives there even though Amazon originally reported through the main MONAI repo.
| Detail | Value (per advisory) |
|---|---|
| Package | monailabel (pip) |
| Affected | 0.8.5 and earlier |
| Patched release | None listed yet; fix merged in PR #1915 for the next release |
| CWE | CWE-22 path traversal |
| Default auth | MONAI_LABEL_AUTH_ENABLE=False — any network client that can reach the server can exploit in default config |
| With auth | ANNOTATOR role is sufficient |
Mechanism (advisory): PUT /datastore/image accepts a caller-controlled image ID joined onto the datastore path without ensuring the resolved path stays inside the datastore. PUT /datastore/label builds destinations the same way. shutil.copy() writes attacker-chosen content. Arbitrary write enables RCE (overwrite app code, authorized_keys, cron) or denial of service.
Workarounds until upgrade:
# Confirm what you have installed
pip show monailabel
# In environment or config for the service:
# MONAI_LABEL_AUTH_ENABLE=True — and do not grant ANNOTATOR to untrusted users
Also: do not expose the server to untrusted networks; run as an unprivileged user with write access limited to the datastore directory.
Amazon states it notified maintainers, who released a fix and advisory, and worked with a government partner to notify victim organizations (non-AWS customers).
Attribution and related Cairn abuse
Amazon assesses moderate confidence for a Chinese-nexus actor based on language, timestamps, and university targeting. Indicators for September 8–10, 2026: 142.4.98.145, 142.4.107.34.
The post cites Gambit Security''s earlier Cairn financial campaign (retail card theft) to show espionage and crime both adopting the same framework—complementing MadPot''s ARTEX memory captures from a different Amazon Threat Intelligence line of work.
What this means if you run AI coding agents
On MONAI Label and research GPUs: Treat every pip install monailabel on a shared lab server like deploying a remote code surface. After you verify version, enable auth, bind to internal networks only, and patch when the release containing #1915 ships. If your agent drafts docker-compose or Kubernetes manifests for imaging pipelines, add a human review gate on volume mounts, host networking, and auth env vars—not just on model weights.
On offensive harnesses in the dev toolchain: Cairn, ARTEX, CyberStrikeAI, and similar projects are one git clone away from a coding agent with shell access. That is not a reason to ban research, but it is a reason to:
- Block unauthenticated binds in workspace policy where you can
- Scan skills and MCP configs before run for curl-to-bash pentest installers
- Log destructive commands (
chmod, masscurl, tunnel listeners)
On defensive tempo: Moses warns that minutes to exploit can beat manual alert review. Pair human oversight with automated containment on high-confidence detections—especially on internet-facing Python app servers common in ML labs.
What Beam can and cannot do here
Beam can flag risky shell commands and destructive edits via Claude Code and Codex PreToolUse hooks, scan skills and MCP configs (11 heuristics plus MCP version-pin check) before you run them, and keep an on-device audit trail with redaction before storage. It does not replace network segmentation on MONAI Label, patch management for monailabel, or kernel isolation on GPU servers. Enforce is hook-based, not a sandbox around every Python API; missing cached policy is fail-open. Beam will not close an exposed Cairn dashboard on port 443—that is still your firewall and auth config.
Concrete checks on research and lab machines
If your team runs MONAI Label for imaging research—common on GPU servers agents provision via IaC—run these checks today:
pip show monailabel
# Note the Version line; 0.8.5 and earlier need the GHSA fix or workarounds
# If you use docker, inspect the image tag and the env the container sets:
docker inspect <container> --format '{{range .Config.Env}}{{println .}}{{end}}' | grep MONAI_LABEL
Confirm MONAI_LABEL_AUTH_ENABLE is not left at default False on any interface reachable beyond a trusted VLAN. Review PUT /datastore/image and PUT /datastore/label access in reverse-proxy logs for path traversal patterns in the image ID or label tag fields (.., absolute paths, dot-prefixed IDs—the advisory says the fix rejects empty or dot-prefixed IDs with HTTP 400).
For Cairn or other offensive harnesses on the same fleet:
# Example local check: listening services that match known agent-framework ports from your runbooks
ss -tlnp
Block inbound access to those listeners at the host firewall even when you "only meant to demo internally." Amazon''s discovery path—public Cairn with full prompt history—is the same mistake NorForge/Reuters reporting highlighted at e-commerce scale.
Rotate model provider keys if any key lived on a MONAI Label host that was internet-facing in September 2026, and review IAM roles attached to GPU nodes for lateral movement after a sitecustomize.py-style persistence event (Python startup hooks are a classic follow-on that both human red teams and agents use).
Why sitecustomize.py matters for Python ML stacks
The Amazon write-up is explicit about mechanism, not just "RCE achieved." sitecustomize.py lives on Python''s site search path and runs automatically when a new interpreter starts (unless disabled with -S). Writing attacker code there turns benign API calls—like triggering a training endpoint that spawns workers—into shell execution without patching the MONAI binary itself.
That pattern shows up wherever coding agents manage Python microservices: a single arbitrary write plus a restart or child process equals code execution under the service user—here root, which is why the advisory stresses unprivileged service accounts as a workaround. Agents that docker exec into containers as root for "quick fixes" recreate the same blast radius.
What developers are saying
The Hacker News thread on retail coverage had no substantive technical comments at the time Amazon published—discussion had not yet accumulated reproduction notes or mitigations worth citing separately.
How this compares to Unit 42''s Cairn report
Readers searching "did Cairn find zero-days" will hit both stories. Unit 42 CL-STA-1238 documents ten confirmed intrusions across nineteen orgs using password spraying, IDOR, and weak APIs—explicitly not novel AI-only exploit classes. Amazon''s November 20 post is narrower and deeper on one medical-imaging stack, with high-confidence zero-day status at exploitation time and a published GHSA. Both are true simultaneously: most agent wins stay boring; the scary tail is autonomous vuln discovery on niche OSS.
For criminal-scale scanning, see also Wiz''s agentic attack fingerprints and Google Mandiant''s ARTEX defense layers.
Frequently asked questions
Did Cairn AI agents find a zero-day on their own?
Amazon Threat Intelligence says yes in the September university campaign: agents tested an upload filename parameter, confirmed path traversal, achieved arbitrary write in ~11 minutes, then sitecustomize.py root shell on the stack later identified as MONAI Label in GHSA-qc22-rmgc-vg94. The flaw had no public advisory at exploitation time; GHSA-qc22-rmgc-vg94 followed on November 18, 2026.
Which MONAI Label versions are affected?
monailabel 0.8.5 and earlier per the advisory. Watch the MONAILabel repo for a release including #1915; until then apply workarounds and restrict exposure.
How do I know if Cairn touched my environment?
You are unlikely to see Cairn on a developer laptop unless someone ran it there. Check internet-facing research apps (MONAI Label, Redis, Qdrant) for Sept 2026 anomalies and IOCs 142.4.98.145 / 142.4.107.34. Hunt exposed agent frameworks using guidance from GitGuardian''s framework census.
Does data leave my machine?
This incident concerns university servers, not Beam telemetry. Coding agents that phone home to model APIs are a separate channel—rotate keys if an agent ever pointed a harness at production.
Is Amazon Threat Intelligence the same team as MadPot ARTEX reporting?
Both posts come from Amazon Threat Intelligence on the AWS Security Blog, but cover different campaigns—ARTEX key theft (November 13) versus Cairn university zero-day (November 20).
Related reading
- Unit 42: Cairn CL-STA-1238 autonomous attacks
- Cairn 1,200-site claims vs verified
- FBI FLASH: ARTEX autonomous penetration testing
- Cert-In CRAFTY FALCON: Cairn and ARTEX in BFSI
Advisory versions, Cairn defaults, and MONAI Label release tags are accurate as of November 20, 2026. Re-check GitHub GHSA-qc22-rmgc-vg94 and the MONAILabel release page before you deploy.

