Amazon Threat Intelligence says a China-linked operator left a public Cairn autonomous pentest server on the internet, pointed agents at East Asian universities with the goal "obtain shell access," and watched them find an unpatched path traversal flaw in MONAI Label in about eleven minutes. In a follow-on session, the agents chained that bug to root on a medical-imaging server by planting a Python startup hook. AWS published the story on its Security Blog on November 20, 2026.
This is not the same October NorForge retail hacking report we covered from Reuters. AWS brings primary telemetry, MONAI Label specificity, and maintainer notification details defenders can act on today.
TL;DR #
| Item | Detail (per AWS Nov 20 post) |
|---|---|
| Tool | Cairn open-source autonomous AI pentest framework |
| Exposure | Operator left Cairn UI without authentication |
| Victim app | MONAI Label medical imaging annotation server |
| Bug | Path traversal on upload endpoints; no CVE at time of exploit |
| Speed | ~11 minutes from API mapping to arbitrary file write |
| Root | sitecustomize.py plus training endpoint spawning Python asroot |
| Advisory | GHSA-qc22-rmgc-vg94 ; fix merged,no release yet |
How AWS discovered the campaign #
On September 10, 2026, Amazon Threat Intelligence spotted a server hosting Cairn without login. Because the instance was public, outsiders could read project configs, actor prompts, and agent output. Prompts were in Simplified Chinese; script timestamps used China Standard Time.
The visible campaign targeted university IP addresses in East Asia outside mainland China for roughly 48 hours starting September 8, 2026. AWS published two indicators of compromise IPv4 addresses (partially redacted in the post) from that window.
Cairn's design is straightforward: you give a starting point and a plain-language objective, and multiple agents plan, execute, and adapt without the operator clicking each step. AWS stresses the same UX that helps defenders also helps attackers who can pay model bills.
When agents fail for hours, then win in minutes #
Circular scan frame with a green highlight spot suggesting automated reconnaissance
AWS highlights two failed long runs to show persistence:
Hardened Redis at one university
Against an unauthenticated Redis instance, agents spent about 16 hours across 148 sequential actions. They tried replication abuse, sandbox escape, config writes, memory corruption, and even Log4Shell attempts against a co-located service. They also used SSRF through a Qdrant instance to map Docker networks and sprayed credentials against PostgreSQL, SSH, and other services. None succeeded.
GPU server credential spraying
Another operation ran 6+ hours on a GPU server. Agents harvested roughly 80 usernames from public faculty pages, papers, and lab sites, then tried 30,000+ password combinations against SSH, an admin dashboard, storage, VPN, and a campus portal. No working login.
Those failures matter: they show agents iterate like junior pentesters with unlimited stamina. The breakthrough came on a soft target: an internet-reachable MONAI Label install with default-style exposure.
The eleven-minute MONAI Label zero-day #
Green lens highlights a spot on a stylized medical imaging ribcage outline
AWS says agents identified the framework, mapped API endpoints, and focused on an upload route with a filename parameter. Path traversal tests succeeded. AWS assesses with high confidence this was a zero-day because no fix, issue, or database entry existed yet for the project.
The GitHub advisory now tracks the flaw:
- Endpoints
PUT /datastore/imageand**PUT /datastore/label** join a caller-controlled ID onto storage paths without keeping writes inside the datastore. MONAI_LABEL_AUTH_ENABLE=Falseby default , so any network client that can reach the port may trigger writes.- Versions 0.8.5 and earlier are affected.
- Fix logic landed in pull request #1915 , merged** October 19, 2026** , butpatched versions: None on the advisory as ofNovember 18, 2026 .
Root shell in a later session
Eleven minutes bought arbitrary file write. Root came later when agents:
- Wrote a reverse shell into**
sitecustomize.py** (auto-imported by Python). - Called a training API endpoint to spawn a fresh Python process.
- Watched that process import the payload and connect back as root .
AWS does not claim patient scans were exfiltrated. It does show machine learning research infrastructure can become a code execution bridge when auth and network controls are loose.
Human operator re-enters after the zero-day #
In another session, AWS says a human stepped in after the agents found the bug. The operator supplied a proxy tunnel into a second university network where the same vulnerable software ran, suggesting manual reuse of the exploit to pivot. Agents then explored additional hosts through that tunnel.
That pattern matches how offensive teams may use agents for breadth and humans for lateral movement once a repeatable exploit exists.
Attribution and parallel Cairn abuse #
AWS attributes the campaign with moderate confidence to a China-nexus actor, citing language, timestamps, and university targeting. It does not map the activity to a named APT.
The post also cites Gambit Security research on a financially motivated actor using Cairn against online retailers, underscoring that Cairn adoption is not limited to espionage. Our earlier Reuters-based Cairn retail story covered different victims but the same framework family.
For a wider agent-misuse index, see explainx.ai's running felony-bench tracker of agents affecting real third parties.
Actions AWS and maintainers took #
AWS says it notified MONAI Label maintainers, who published a fix path and GitHub advisory. Victim universities were not AWS customers, so AWS worked through a government partner to alert them.
Behind the scenes, AWS highlights honeypots and network-level restrictions on scanning activity. That is infrastructure-level defense, not a substitute for patching research software you expose.
Defender playbook for ML research stacks #
Shield with a hairline crack under a green security patch
Network and auth defaults
MONAI Label is built for trusted lab networks, not open internet edges. Until a patched package ships:
- Set
MONAI_LABEL_AUTH_ENABLE=Trueand limitANNOTATOR roles. - Block public ingress; require VPN or private links.
- Run the service account with minimal write permissions outside the datastore.
Speed assumptions
AWS argues SOC timelines tuned for human attackers lag autonomous agents that never . Pair high-confidence detections with automated containment, with humans reviewing outcomes instead of every single tool call.
Coding agents that can hit ML admin APIs need outbound controls. AgentBeam can block dangerous agent actions before they reach exposed research endpoints.
Patch tracking
Watch GHSA-qc22-rmgc-vg94 and the MONAI Label release page. A merged fix on main is not the same as a version your orchestrator can pull.
How Cairn compares to other agent offense stories #
| Incident | What differed |
|---|---|
| [NorForge retail campaign (Oct)](https://explainx.ai/blog/hacker-ai-agents-5000-dollars-deere-sharkninja-cummins-breach-norforge-cairn-2026) | Journalist-sourced, e-commerce focus, mixed models |
| AWS MONAI case (Nov) | Primary AWS blog, medical ML stack, measured 11-minute zero-day window |
| OpenAI CI root misalignment report | Benign eval agents misusing peer credentials |
What we cannot verify from AWS alone #
AWS describes its reconstruction from a public Cairn instance. Independent media have not re-tested the MONAI exploit chain. Clinical impact (PHI exposure) is not claimed.
Treat root on one server as a warning for any default-open MONAI Label on your perimeter, not proof every install falls the same way.
What is CVE-2026-133968? #
Outlets including The Hacker News report that the MONAI Label flaw now carries CVE-2026-133968 with a 9.8 critical CVSS score. The GitHub advisory timeline still shows patched versions: None while maintainers prepare a release. Treat the CVE as the tracking ID for scanners even if your package manager cannot upgrade yet.
Workarounds from the advisory
Until you can install a fixed release, MONAI Label's advisory recommends:
- Enabling authentication and withholding ANNOTATOR from untrusted users.
- Keeping the server off public networks.
- Running the process as an unprivileged user with writes limited to the datastore tree.
Those controls would not stop a determined insider, but they block the anonymous network client path AWS describes.
Why Redis and MONAI sat on the same campaign #
The Redis stretch matters for defenders who think "we patched Redis years ago." AWS documents 148 distinct attempts against a misconfigured instance. Agents treated failure as signal, not stop. Only when they found a web-facing ML admin API with permissive uploads did the timeline collapse to minutes.
If your org runs research GPUs and legacy data stores on overlapping subnets, segment them. Agents chaining SSRF from Qdrant to internal PostgreSQL shows how graph-like exploration crosses service boundaries humans forget to firewall.
Red-team versus real attacker economics #
Cairn's README celebrates defender use in hackathons. AWS and Gambit show the same UI on a VPS with a stolen API key budget. Your red team using Cairn with authorization is legitimate; an exposed Cairn dashboard without auth is a live operation journal for threat hunters.
Document that difference in policy: pentest agents belong in isolated VPCs, not port 443 on a default cloud image.
Clinical and research data boundaries #
MONAI Label often sits near imaging pipelines that are not HIPAA production systems but still hold identifiable research scans. AWS does not claim imaging theft in this post. The lesson is architectural: labeling servers are production-adjacent and attract agents scanning university IP space.
Pair this incident with ECHR lawyer sanctions for AI-hallucinated citations as another reason research IT cannot treat ML tooling as "lab only" without network controls.