Overview
In a landmark incident reported by Amazon Threat Intelligence, suspected China-linked threat actors deployed autonomous AI agents to discover and exploit a previously unknown path-traversal vulnerability in MONAI Label, an open-source medical imaging annotation server widely used in academic and healthcare research environments. The agents achieved file-write capability within roughly 11 minutes and later escalated to a full root shell on at least one university server. A second university network is believed to have been compromised via the same method. No patched release of MONAI Label is currently available, though maintainers merged a source-level fix on October 19, 2026.
Technical Analysis
The vulnerability is a path-traversal flaw in MONAI Label’s file upload handlers. File identifiers supplied by callers are appended directly to the server’s storage directory path without sanitisation, allowing an attacker to write files to arbitrary writable locations on the host filesystem. Because MONAI Label runs with authentication disabled by default, any network-reachable client can trigger the write without credentials.
Root access was achieved in a subsequent session via a two-stage technique:
- The agents used the path-traversal flaw to plant a malicious file in a location Python loads automatically at startup (e.g., a
.pthfile or sitecustomize module). - They then invoked MONAI Label’s training feature, which spawned a new Python process. That process loaded the planted file and established a reverse shell back to the attacker with root privileges.
Amazon reconstructed the attack chain because the operators inadvertently left their Cairn control panel — an open-source AI agent orchestration framework — exposed on the internet without authentication, revealing task prompts, agent run histories, and configuration in plaintext. Forensic indicators including Simplified Chinese-language prompts and China Standard Time process timestamps informed the moderate-confidence attribution.
Framework Mapping
- AML.T0103 (Deploy AI Agent): Attackers operationalised autonomous AI agents as offensive penetration tools with goal-directed behaviour.
- AML.T0084 (Discover AI Agent Configuration): Amazon’s reconstruction was enabled by the exposed Cairn panel, itself a demonstration of how AI agent configuration leakage creates bilateral risk.
- AML.T0047 (AI-Enabled Product or Service): MONAI Label, as an AI-serving infrastructure component, was the targeted attack surface.
- LLM08 (Excessive Agency): The Cairn agents operated with broad autonomy — selecting attack methods and chaining exploitation steps without per-action human approval.
- LLM07 (Insecure Plugin Design): The file-upload handler functioned as an insecure tool/plugin with no input boundary enforcement.
Impact Assessment
The immediate impact is confirmed root-level compromise of at least one university research server and suspected compromise of a second. The broader implication is significant: this appears to be one of the first publicly documented cases of autonomous AI agents being used in a live, targeted intrusion to independently discover and exploit a novel vulnerability. All MONAI Label deployments on versions 0.8.5 and earlier — particularly those exposed to any network without authentication — remain unpatched and vulnerable.
Mitigation & Recommendations
- Network isolation: Immediately place MONAI Label servers behind a VPN or restrict access to trusted IP ranges using firewall rules.
- Enable authentication: Override the default no-auth configuration wherever possible pending an official patched release.
- Audit startup paths: Check Python startup file locations (site-packages,
.pthfiles, sitecustomize) for unexpected additions. - Monitor Cairn and similar frameworks: Ensure any AI agent orchestration panels are never exposed publicly; require strong authentication on all management interfaces.
- Track upstream releases: Watch the MONAI Label repository for an official versioned release incorporating the October 19 source fix before upgrading.