LIVE FEED
ATLAS OWASP CRITICAL Active exploitation · Immediate action required RELEVANCE ▲ 9.2

MONAI Label Zero-Day Exploited by China-Linked AI Agents

TL;DR CRITICAL
  • What happened: China-linked AI agents autonomously found and exploited a zero-day in MONAI Label medical software.
  • Who's at risk: Universities and healthcare institutions running MONAI Label 0.8.5 or earlier with default unauthenticated network access are directly exposed.
  • Act now: Immediately restrict network access to MONAI Label instances and enable authentication · Upgrade to the patched MONAI Label release as soon as it is published · Audit exposed AI pentesting frameworks like Cairn for unauthenticated access
MONAI Label Zero-Day Exploited by China-Linked AI Agents

Overview

In September 2026, Amazon’s threat intelligence team identified a China-linked campaign in which autonomous AI agents exploited a previously unknown vulnerability in MONAI Label, an open-source tool for AI-assisted annotation of medical images. The agents operated via Cairn, an open-source AI penetration testing framework, and conducted an end-to-end attack cycle with minimal human oversight. The incident marks a qualitative shift in offensive AI capability: autonomous agents, not human operators, performed reconnaissance, vulnerability discovery, and exploitation.

Technical Analysis

Amazon discovered an exposed, unauthenticated Cairn instance that allowed full visibility into operator prompts and agent outputs. The operator provided agents with a list of East Asian university targets and a goal of obtaining shell access.

The agents followed a structured kill chain:

  1. API Mapping — Agents identified the software stack on a target server and mapped its API surface.
  2. Vulnerability Discovery — Within the API, agents identified a path traversal bug in a file upload function.
  3. Initial Exploitation — Approximately 11 minutes after beginning API mapping, agents were arbitrarily writing files to the server.
  4. Persistence & Privilege Escalation — In a follow-on session, agents planted a Python reverse shell inside a file Python executes on interpreter startup, achieving a root shell.
  5. Lateral Movement — The operator supplied a proxy tunnelling into a second university. Amazon inferred this pivot host had also been compromised via the same zero-day.

The vulnerability affects MONAI Label versions 0.8.5 and earlier. Authentication is disabled by default, meaning any network-reachable instance is exposed. The MONAI Label project published a critical advisory coinciding with Amazon’s disclosure; a fix will be included in the next release.

# Simplified conceptual representation of the path traversal vector
POST /api/upload?filename=../../startup/sitecustomize.py
Content: 

Framework Mapping

  • AML.T0103 (Deploy AI Agent): The threat actor deployed Cairn-based autonomous agents to conduct the entire attack lifecycle.
  • AML.T0084 (Discover AI Agent Configuration): The exposed Cairn instance leaked operator prompts and agent outputs, illustrating risks from misconfigured AI tooling.
  • AML.T0047 (AI-Enabled Product or Service): MONAI Label itself is an AI-assisted medical tool, making this an attack on AI infrastructure within healthcare.
  • LLM08 (Excessive Agency): The agents operated autonomously from reconnaissance through exploitation and persistence with minimal human checkpoints.
  • LLM07 (Insecure Plugin Design): The file upload API in MONAI Label lacked adequate path sanitisation, a classic insecure tool/plugin pattern.

Impact Assessment

The immediate victims are universities in East Asia running unpatched MONAI Label instances. Healthcare and research institutions globally using MONAI Label 0.8.5 or earlier with default settings are at risk. The broader implication is strategic: this is one of the first publicly documented cases of autonomous AI agents conducting zero-day discovery and exploitation in a live campaign, lowering the bar for sophisticated intrusions against under-resourced targets such as academic medical centres.

Mitigation & Recommendations

  1. Restrict network exposure — Place MONAI Label behind a firewall or VPN; it should never be publicly internet-accessible.
  2. Enable authentication — Override the default unauthenticated configuration immediately.
  3. Patch promptly — Apply the forthcoming MONAI Label fix as soon as it is released.
  4. Audit AI pentesting tools — Ensure frameworks like Cairn are never deployed without authentication; exposed instances leak operator intent and agent capabilities.
  5. Monitor for path traversal patterns — Deploy WAF rules and file integrity monitoring targeting unusual writes to Python startup paths.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.