MONAI Label CVE-2026-133968: datastore path traversal found and exploited as a zero-day by Chinese-nexus Cairn AI agents for root, no fixed release yet
SockAnalyst · Security+ certified analyst.
What happened
Amazon Threat Intelligence reported this week that a suspected Chinese-nexus actor used Cairn, an open-source autonomous AI penetration-testing framework, to find and exploit a previously unknown flaw in MONAI Label, the open-source server research teams use to label medical scans and train imaging models. Amazon found the actor's Cairn instance exposed without authentication on September 10, 2026; its history showed a roughly 48-hour campaign from September 8 against university IP addresses in East Asia outside mainland China, with prompts in Simplified Chinese and goals such as 'obtain shell access'. Against a MONAI Label server, the agents mapped the API and in about 11 minutes went from that to writing arbitrary files through an upload endpoint's path traversal. In a later session they wrote a Python reverse shell into sitecustomize.py (which Python imports at every interpreter start), triggered a training API endpoint to spawn a new Python process, and got a reverse shell running as root. The operator then gave the agents a proxy into a second university network running the same software; Amazon infers, but has not confirmed, that it was compromised the same way. Other attempts failed, including a 16-hour, 148-action effort against a hardened Redis server. Amazon attributes the activity to a Chinese-nexus actor with moderate confidence. The maintainers' advisory GHSA-qc22-rmgc-vg94 (Nov 18, 2026) says PUT /datastore/image takes a caller-controlled image parameter as the file ID without validation and copies the upload to the resulting path; PUT /datastore/label builds its path the same way. Authentication is off by default (MONAI_LABEL_AUTH_ENABLE=False), so any client that can reach the server can exploit it; with authentication on, the ANNOTATOR role is enough. MITRE published CVE-2026-133968 (CVSS 3.1 9.8) on November 20. The fix is merged in source (PR #1915) but the advisory lists no patched release; 0.8.5 is still the newest release.
Who's affected
Research hospitals, universities and AI/ML teams running MONAI Label 0.8.5 or earlier, especially servers reachable from the internet or shared research networks, running in the default no-authentication setup, or running as root. MONAI Label is a separate program from the broader MONAI framework; MONAI-only installs are not named in the advisory. Hosts that also hold medical images, model training data, GPU credentials or SSH keys are the most valuable targets.
✅ Fix / patch
- There is no fixed release yet. Until one ships, follow the advisory's workarounds: enable authentication (MONAI_LABEL_AUTH_ENABLE=True), do not give the ANNOTATOR role to untrusted users, and do not expose MONAI Label to untrusted networks (take internet-facing instances offline or put them behind a VPN).
- Run MONAI Label as an unprivileged service account that can write only to its datastore directory, never as root, so a traversal write cannot reach site-packages, ~/.ssh/authorized_keys or cron.
- If you build from source, pick up the merged fix in PR #1915 (paths are resolved and rejected if they leave the datastore image/label directories; empty or dot-prefixed IDs are refused) and move to the next official release when it ships.
- Check for compromise: look for sitecustomize.py or usercustomize.py files you did not create in every Python environment on the host, files written outside the datastore, new SSH keys or cron entries, and outbound shells from Python processes. Block and search for the two Amazon-published IPs. If found, isolate the host, rebuild it, and rotate any credentials, tokens and SSH keys it held.
🔎 Quick hunt check
Amazon published two IPs seen September 8-10, 2026 (below); no hashes or domains. Behaviour to hunt: (1) PUT requests to /datastore/image or /datastore/label whose image or tag parameter contains ../, ..%2f or %2e%2e (MONAI Label listens on port 8000 by default); (2) creation or modification of sitecustomize.py / usercustomize.py in any Python site-packages directory; (3) Python processes on the MONAI Label host spawning shells or opening outbound connections to the internet, especially right after a training API call; (4) traffic to or from 142.4.98[.]145 and 142.4.107[.]34.
MITRE ATT&CK: T1190, T1546.018, T1059.006, T1090, T1110
IOCs
- IP addresses:
142.4.98[.]145,142.4.107[.]34 - Files / paths:
sitecustomize.py written by the attacker's agents (Python reverse shell loaded at interpreter start)
IPs published by Amazon Threat Intelligence, from records dated Sept 8-10, 2026. No hashes or domains published. T1110 reflects the campaign's failed credential brute-forcing against another university server, not the MONAI Label exploit.
Hunt queries
Starting points built only from published indicators. Field names vary; tune before use.
```MONAI Label CVE-2026-133968: traversal in /datastore uploads, sitecustomize.py writes, Python shells, Amazon-published IPs```
index=* earliest=09/01/2026:00:00:00
("/datastore/image" OR "/datastore/label" OR "sitecustomize.py" OR "usercustomize.py" OR "142.4.98.145" OR "142.4.107.34" OR parent_process_name IN ("python","python3") OR ParentImage="*python*")
| eval req=urldecode(coalesce(uri, url, uri_query, "")), method=upper(coalesce(http_method, method, "")), path=coalesce(file_path, TargetFilename, file_name, ""), child=lower(coalesce(process_name, Image, "")), parent=lower(coalesce(parent_process_name, ParentImage, ""))
| eval signal=case(
match(req,"/datastore/(image|label)") AND match(req,"(\.\./|\.\.\\\\)"), "datastore_path_traversal",
match(path,"(?i)(site|user)customize\.py$"), "python_startup_hook_written",
match(_raw,"142\.4\.98\.145|142\.4\.107\.34"), "amazon_ioc_ip",
match(parent,"python[0-9.]*$") AND match(child,"(^|/)(sh|bash|dash|nc|ncat|socat)$"), "python_spawned_shell",
true(), null())
| where isnotnull(signal)
| stats count min(_time) as first_seen max(_time) as last_seen values(signal) as signals values(req) as requests values(path) as files values(child) as children by host, src_ip
| convert ctime(first_seen) ctime(last_seen)
| sort - count// MONAI Label CVE-2026-133968: /datastore traversal, sitecustomize.py writes, Python-spawned shells, Amazon-published IPs
let Since = datetime(2026-09-01);
let AmazonIPs = dynamic(["142.4.98.145","142.4.107.34"]);
union isfuzzy=true
(CommonSecurityLog | where TimeGenerated >= Since
| where (RequestURL has_any ("/datastore/image","/datastore/label") and url_decode(RequestURL) has "../")
or SourceIP in (AmazonIPs) or DestinationIP in (AmazonIPs)
| extend Signal = iff(SourceIP in (AmazonIPs) or DestinationIP in (AmazonIPs), "amazon_ioc_ip", "datastore_path_traversal"), Who = coalesce(DestinationHostName, DestinationIP), Evidence = coalesce(RequestURL, strcat(SourceIP, " -> ", DestinationIP))),
(DeviceFileEvents | where TimeGenerated >= Since
| where FileName in~ ("sitecustomize.py","usercustomize.py") and ActionType in ("FileCreated","FileModified")
| extend Signal = "python_startup_hook_written", Who = DeviceName, Evidence = strcat(FolderPath, " by ", InitiatingProcessCommandLine)),
(DeviceProcessEvents | where TimeGenerated >= Since
| where InitiatingProcessFileName startswith "python" and FileName in~ ("sh","bash","dash","nc","ncat","socat")
| extend Signal = "python_spawned_shell", Who = DeviceName, Evidence = ProcessCommandLine),
(DeviceNetworkEvents | where TimeGenerated >= Since
| where RemoteIP in (AmazonIPs)
| extend Signal = "amazon_ioc_ip", Who = DeviceName, Evidence = strcat(RemoteIP, ":", tostring(RemotePort), " ", InitiatingProcessFileName))
| summarize Hits = count(), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated), Signals = make_set(Signal), Evidence = make_set(Evidence, 20) by Who
| order by Hits desc⏱ CISA deadline
Not in CISA KEV as of Nov 22, 2026.
Sources
- Project MONAI: GHSA-qc22-rmgc-vg94, Path traversal in datastore upload endpoints allows arbitrary file write (Nov 18, 2026)
- Amazon Threat Intelligence (AWS Security Blog): Chinese-nexus actor's Cairn AI agents found a zero-day on their own
- CVE record: CVE-2026-133968 (published Nov 20, 2026)
- Project-MONAI/MONAILabel PR #1915 (fix)
- BleepingComputer: Hackers use autonomous AI tool Cairn to find zero-day, take over server (Nov 20, 2026)
- The Hacker News: Attackers' AI Agents Found a Zero-Day in Medical-Imaging Software and Gained Root Access (Nov 21, 2026)
Related playbooks
- Cisco Expressway CVE-2026-76466: clustered database-service root RCE now exploited after public PoC, four nodes breached · Exploited vuln · 2026-11-20
- LiteLLM CVE-2026-40217: guardrail test endpoint sandbox escape runs code as root, now exploited (CISA KEV) · Exploited vuln · 2026-11-19
- SAP NetWeaver Enterprise Portal CVE-2025-42884: year-old unauthenticated JNDI injection now exploited, CISA KEV deadline Nov 23 · Exploited vuln · 2026-11-21
Get the SockAnalyst Weekly Brief free, every Thursday.
One email a week with the KEV deadlines, patches and hunt queries that matter. New playbooks still post here on the site every morning.
Subscribe free
SockAnalyst