cd /news/ai-safety/cairn-autonomously-discovers-monai-l… · home › topics › ai-safety › article
[ARTICLE · art-173592] src=dev.to ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Cairn Autonomously Discovers MONAI Label Zero-Day and Gains Root Shell

AWS Security reported that an autonomous AI agent tool called Cairn, assessed with medium confidence as operated by a China-nexus actor, independently discovered a previously unannounced path-traversal vulnerability in MONAI Label 0.8.5 and earlier, moving from arbitrary file write to a root-privileged reverse shell in roughly 11 minutes. The agent enumerated unauthenticated datastore APIs, wrote a reverse shell into sitecustomize.py via the image parameter, and triggered Python execution through the training API. AWS said the China-nexus attribution does not imply Chinese government involvement, and follow-on compromise of a separate university network remains an inference rather than a confirmed finding.

recommendation

Require authentication on MONAI Label datastore and training APIs, run the service as an unprivileged user with minimal write permissions, and block outbound shell traffic to prevent agent-driven file-write-to-RCE chains.

more in playbook: Securing agents and AI tools
by read6 min views14 publishedNov 21, 2026

#

  1. Basic Information

Article Name: Chinese-nexus actor's Cairn AI agents found a zero-day on their own #

Publisher: AWS Security #

Publication Date: 2026-11-20 #

Collection Date: 2026-11-21T07:45:23+09:00 #

Source:Primary Source #

Severity: Critical (Author's triage priority) #

Reason for Severity: Confirmed in real environments from the discovery of an unannounced vulnerability to arbitrary file write and root shell acquisition via Python execution. The China-nexus assessment does not imply Chinese government involvement.

#

  1. Executive Summary

AWS reported an incident in which an attacker assessed with medium confidence as China-nexus instructed Cairn to obtain a shell; Cairn discovered a MONAI Label path traversal vulnerability in about 11 minutes and progressed from arbitrary file write to a root shell.

#

  1. Attack Flow

From MONAI Label Zero-Day to Root Shell

  1. Cairn enumerates APIs and explores endpoints such as PUT /datastore/image.
  2. Writes files outside the datastore using path traversal in the image parameter.
  3. Places a reverse shell into sitecustomize.py.
  4. Launches Python via the training API to obtain a root shell.

#

  1. Attacker Location and Execution Environment
  • A remote attacker who can reach unauthenticated MONAI Label APIs from the internet.

#

  1. Victim and Administrator Perspective

Victim

  • Host files and processes may change even while the service appears to operate normally.

Administrator

  • Visible as anomalous requests to the datastore API, Python execution, outbound shell traffic, and modifications to sitecustomize.py.

#

  1. Conditions for Success and Failure

Success Conditions

  • MONAI Label 0.8.5 or earlier
  • Attacker access to PUT /datastore/image or save_label paths
  • Writable paths for the service account and a subsequent Python execution pathway

Failure and Mitigation Conditions

  • Enable authentication and restrict ANNOTATOR to trusted users
  • Block the service from untrusted networks
  • Run with an unprivileged user and minimal write permissions
  • Apply the fixed release from the maintainer

#

  1. What Happens Upon Success
  • Arbitrary file write on the MONAI Label host
  • Code execution under the privileges of the MONAI Label process (root-privileged reverse shell in AWS's observed cases)
  • Potential subsequent compromise of the same network and credentials (unconfirmed in the observed cases of this article)

#

  1. Observable Logs

Email: None noted #

Proxy / SWG / DNS: Enumeration and outbound callbacks to the MONAI Label API from Cairn workers or unknown sources #

Endpoint / EDR: Modification of Python startup files such as sitecustomize.py / python child process following the training API and shell connection #

Identity / IdP: None noted #

SaaS / Cloud: None noted #

Network: Short-term multi-endpoint access to the MONAI Label API / reverse-shell traffic from the server to an unknown host

#

  1. Attack Success Assessment

Initial Execution Confirmed: Public information: AWS reported from operational history of public Cairn instances that, in a separate session after discovering the arbitrary file write, a root-privileged reverse shell was established via Python execution through the training API. Compromise of a separate university network and tunnel establishment are AWS inferences and are not treated as confirmed follow-on compromise facts. Scope: MONAI Label servers observed by AWS.

Trials, executions, privilege acquisition, information access, and follow-on compromises are separated; a single IOC or request does not determine attack success.

#

  1. Investigation Playbook

Trigger: Anomalous enumeration of the MONAI Label API or file writes outside the datastore #

Initial Checks: Check version, authentication, external reachability, and service privileges / preserve API, file, process, and network logs #

Endpoint: Check Python startup files such as sitecustomize.py, training processes, and shell history #

Identity / Cloud: Check service accounts, tokens, and research data access #

Follow-on: Track proxying to a second host, credential usage, and data access #

Containment: Isolate the service, enable authentication, reduce privileges, and rotate credentials #

Classification: Separate reconnaissance, arbitrary file write, Python execution, root shell, and lateral movement

#

  1. Defense and Detection Ideas

Single Event: Image parameters containing paths outside the datastore / unexpected modifications to sitecustomize.py #

Timeline: Correlate API enumeration -> file write -> training API -> outbound shell traffic #

Hunting: Cross-search for Python startup files and outbound connections across MONAI Label host groups #

Log Gaps: Check for missing API bodies, file audits, process trees, and egress data #

Priority Mitigations: Block external access, enable authentication, enforce least privilege, and track fixed releases

#

  1. Facts, Inference, and Hypothesis

Facts

  • AWS assessed the attacker as China-nexus with medium confidence based on Simplified Chinese prompts, script timestamps configured to China Standard Time, and target selection. This does not directly indicate Chinese government direction or involvement.
  • The attacks occurred between September 8 and 10, 2026, primarily targeting universities in East Asia outside mainland China.
  • Cairn explored MONAI Label APIs and discovered an arbitrary file write path using the image parameter in PUT /datastore/image in about 11 minutes.
  • The attack session wrote a Python reverse shell into sitecustomize.py and launched Python via the training API to obtain a root-privileged shell.
  • The GHSA lists MONAI Label 0.8.5 and earlier as affected, with authentication disabled by default. Even when authentication is enabled, the ANNOTATOR role can perform the operation.

Inference

  • Proxying to another university network with the same vulnerable software indicates that the network may have already been compromised, but this is an AWS estimate and not independently confirmed by the victim organizations.

Hypothesis

  • None. Damage scope or attack success beyond public information is not asserted.

#

  1. MITRE ATT&CK Mapping

| ID | Technique | Confidence | Basis | | T1190 | Exploit Public-Facing Application | high | Utilized path traversal in the public MONAI Label API. | | T1105 | Ingress Tool Transfer | high | Wrote reverse shell code into sitecustomize.py. | | T1059.006 | Python | high | Launched Python from the training API to execute a reverse shell. |

#

  1. Unknowns and Further Investigation
  • Release timeline for the patch by MONAI Label maintainers
  • Scope of compromise and independent confirmation in the second university network

#

  1. Impact on SOCs and Organizations

In medical imaging AI and research environments, prioritize checking for unauthenticated public MONAI Label instances, container or root execution, and child processes spawned from training APIs.

#

  1. Audience Summary

SOC: Correlate exploration of MONAI Label APIs, file writes outside the datastore, and outbound shell traffic immediately following training API calls. #

Administrators: Enable authentication, block access from untrusted networks, run with unprivileged users and restricted write permissions, and track fixed releases. #

Users: Check whether research servers are publicly exposed, and have administrators investigate possible compromise; end-user actions alone cannot prevent exploitation.

── more in #ai-safety 4 stories · sorted by recency
── more on @aws security 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/cairn-autonomously-d…] indexed:0 read:6min 2026-11-21 · —