Executive Summary
In September 2026, suspected China-linked threat actors deployed autonomous AI agents via the Cairn open-source penetration-testing framework to discover and exploit CVE-2026-133968, a critical (CVSS 9.8) path-traversal zero-day in MONAI Label 0.8.5, an open-source medical-imaging annotation server. Within approximately 11 minutes of automated reconnaissance, the AI agents identified the unauthenticated file-write vulnerability, subsequently planting a malicious Python startup file that—when triggered through the server's training feature—returned a root shell to the attackers. At least one university research server was confirmed compromised, and a second institution is under investigation. No fixed installable release existed as of the advisory publication date, leaving all MONAI Label deployments on version 0.8.5 or earlier exposed.
This incident marks a documented shift in the threat landscape: adversaries are now operationalizing agentic AI to autonomously discover zero-days at machine speed, compressing the window between vulnerability existence and active exploitation to under an hour. The convergence of AI-augmented offense, unauthenticated-by-default medical research software, and healthcare-adjacent infrastructure signals an accelerating threat curve that security teams and compliance officers can no longer treat as theoretical.
Why This Matters Now
Autonomous AI agents discovering and exploiting critical zero-days in under 11 minutes fundamentally breaks traditional patch-and-detect timelines. With CVE-2026-133968 rated 9.8 and no fixed package available, every reachable MONAI Label deployment remains live bait. The healthcare and research sectors must immediately enforce network segmentation, authentication, and least-privilege controls—not wait for a vendor release.
Attack Path Analysis
China-linked threat actors deployed autonomous AI agents via the Cairn framework to discover and exploit CVE-2026-133968, a critical path-traversal zero-day in MONAI Label medical imaging servers running with no authentication, writing a malicious Python startup file to gain root shell access on a university server. The agents operated with minimal human direction over 48 hours, escalating from unauthenticated file writes to root code execution, then pivoting into a second university network via an operator-supplied tunnel, with the entire operation observable only after the attackers inadvertently exposed their own Cairn control panel without a password.
Kill Chain Progression
This analysis maps confirmed threat intelligence to the full cloud kill chain to show where defensive gaps would emerge as an attack progresses.
Initial Compromise
Description
AI agents operating under the Cairn autonomous penetration testing framework discovered and exploited CVE-2026-133968, a critical (CVSS 9.8) path-traversal vulnerability in MONAI Label 0.8.5. The server's default configuration required no authentication, allowing any network-reachable client to submit malicious file identifiers that wrote files outside the designated storage folder. The agents autonomously mapped the application interface and reached the file-write primitive in approximately 11 minutes without prior knowledge of the flaw.
Related CVEs
CVE-2026-133968
CVSS 9.8A path-traversal vulnerability in MONAI Label 0.8.5 and earlier allows an unauthenticated remote attacker to write arbitrary files outside the designated storage folder due to insufficient sanitization of file identifiers supplied by the caller, potentially leading to remote code execution with the privileges of the server process including root access.
Affected Products:
Project MONAI MONAI Label – <= 0.8.5
Exploit Status:
exploited in the wildReferences:
https://www.cve.org/CVERecord?id=CVE-2026-133968https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94https://github.com/Project-MONAI/MONAILabel/pull/1915https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/https://pypi.org/project/monailabel/
MITRE ATT&CK® Techniques
Exploit Public-Facing Application
Command and Scripting Interpreter: Python
Hijack Execution Flow: COR_PROFILER
Event Triggered Execution: Installer Packages
Abuse Elevation Control Mechanism
File and Directory Discovery
Ingress Tool Transfer
Exploitation for Client Execution
Potential Compliance Exposure
Mapping incident impact across multiple compliance frameworks.
HIPAA Security Rule (45 CFR Part 164) – Access Control and Integrity Controls for ePHI Systems
Control ID: 164.312(a)(1) / 164.312(c)(1)
NIST SP 800-53 Rev. 5 – Information Input Validation / Access Enforcement / Least Functionality
Control ID: SI-10 / AC-3 / CM-7
NIS2 Directive (EU) 2022/2555 – Risk Management and Vulnerability Handling for Essential Entities
Control ID: Article 21(2)(a) / Article 21(2)(e)
CISA Zero Trust Maturity Model 2.0 – Network Segmentation and Identity-Centric Access Enforcement
Control ID: Network Pillar - Macro/Micro Segmentation / Identity Pillar - Authentication
ISO/IEC 27001:2022 – Management of Technical Vulnerabilities / Access to Source Code / Identity and Access Management
Control ID: Annex A 8.8 / Annex A 8.4 / Annex A 5.15
DORA (EU Regulation 2022/2554) – ICT Security Controls, Vulnerability Management, and Incident Detection
Control ID: Article 9(2) / Article 10(1) / Article 13(2)
Sector Implications
Industry-specific impact of the vulnerabilities, including operational, regulatory, and cloud security risks.
Health Care / Life Sciences
AI-augmented zero-day exploitation of MONAI Label medical imaging servers exposes unpatched CVE-2026-133968 (CVSS 9.8), threatening root-level compromise of clinical imaging infrastructure and violating HIPAA.164.312 data-in-transit protections.
Higher Education/Acadamia
China-linked AI agents gained root access on university MONAI Label servers via path-traversal, exploiting default no-auth configurations and demanding immediate Zero Trust segmentation and East-West traffic controls per NIST.800-53.SC-7.
Research Industry
Autonomous Cairn AI agents autonomously discovered and weaponized a zero-day in medical imaging research software, threatening intellectual property exfiltration and requiring urgent egress policy enforcement and multicloud visibility controls.
Information Technology/IT
AI-augmented intrusion via open-source MONAI Label exposes IT teams to agentic AI-driven lateral movement and privilege escalation, demanding Cloud Native Security Fabric deployment and inline IPS controls mapped to NIST.CSF.DE.CM-4.
Sources
- Attackers' AI Agents Found a Zero-Day in Medical-Imaging Software and Gained Root Accesshttps://thehackernews.com/2026/11/attackers-ai-agents-found-zero-day-in.htmlVerified
- Chinese-Nexus Actors: Cairn AI Agents Found a Zero-Day on Their Own (AWS Security Blog)https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/Verified
- MONAI Label Security Advisory GHSA-qc22-rmgc-vg94 (GitHub)https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94Verified
- MONAI Label Path-Traversal Fix Pull Request #1915 (GitHub)https://github.com/Project-MONAI/MONAILabel/pull/1915Verified
- CVE-2026-133968 Record (CVE.org)https://www.cve.org/CVERecord?id=CVE-2026-133968Verified
- MONAI Label PyPI Release Historyhttps://pypi.org/project/monailabel/Verified
Frequently Asked Questions
Cloud Native Security Fabric Mitigations and ControlsCNSF
Based on the attack progression modeled above, these are the defensive controls that would constrain each stage.
Aviatrix Zero Trust CNSF is directly relevant to this incident because the MONAI Label server was openly reachable from the network with no segmentation boundary controlling inbound access, allowing AI-driven agents to exploit it within minutes. By enforcing workload-level segmentation and restricting east-west and egress paths, Aviatrix would likely have constrained the attacker's reachability and reduced the blast radius across both university environments.
Control: Cloud Native Security Fabric (CNSF)
Mitigation: Restricting which network sources could reach the MONAI Label service would likely have reduced the attack surface available to the AI agents, making the unauthenticated endpoint considerably harder to reach from arbitrary external or lateral positions.
Control: Zero Trust Segmentation
Mitigation: Zero Trust segmentation would likely constrain the scope of what the compromised MONAI Label workload could reach once the malicious process executed, limiting the outbound connection paths available to the resulting root shell and reducing its operational value to the attacker.
Control: East-West Traffic Security
Mitigation: East-west traffic enforcement would likely reduce the attacker's ability to pivot across institutional network boundaries via operator-supplied tunnels, constraining lateral reachability between the two university environments even when using pre-established connectivity.
Control: Multicloud Visibility & Control
Mitigation: Centralized visibility across the university's cloud workloads would likely surface the anomalous outbound callback traffic from the MONAI Label server, and policy enforcement at the fabric level could reduce the persistence of the command channel by flagging or constraining unauthorized outbound connections to unknown external infrastructure.
Control: Egress Security & Policy Enforcement
Mitigation: Egress policy enforcement would likely constrain the outbound data transfer paths available to the attacker from the compromised MONAI Label server, reducing the volume and ease of any potential data movement toward external attacker-controlled destinations.
Had prior segmentation and egress controls been in place across both university environments, the attacker's confirmed root-level foothold would likely have remained more isolated, with reduced ability to extend persistence across networks or reach sensitive medical imaging datasets freely.
Impact at a Glance
Affected Business Functions
- Medical Imaging Annotation and AI Model Training Pipelines
- Research Data Integrity and Clinical Research Infrastructure
- University Network Security and Server Administration
- Healthcare Research Data Governance and Compliance
Estimated downtime: 7 days
Estimated loss: N/A
Medical imaging datasets (radiology and pathology scans used for AI training), AI model training configurations and proprietary research data hosted on university MONAI Label servers. No confirmed exfiltration of patient health records or PII was established; however, root-level access on at least one university server means all data accessible to the server process was at risk. A second university network is suspected compromised, compounding exposure. The unauthenticated write primitive combined with root shell access creates material risk of unauthorized access to research intellectual property, model weights, and potentially patient-linked scan metadata depending on institutional data practices.
Recommended Actions
Key Takeaways & Next Steps
- • Apply Cloud Firewall (ACF) and Egress Security controls to restrict MONAI Label servers to allow-listed inbound sources only, blocking unauthenticated network access to research imaging servers at the network layer independent of application-level authentication settings — directly addressing the default no-auth exposure that enabled the zero-day exploit.
- • Deploy Zero Trust Segmentation with identity-based microsegmentation policies to enforce least-privilege workload-to-workload boundaries, ensuring that a compromised MONAI Label server cannot serve as a pivot point into adjacent university networks or other research infrastructure, limiting the east-west lateral movement observed between university targets.
- • Enable Multicloud Visibility and Threat Detection controls to baseline normal MONAI Label API traffic patterns and generate alerts on anomalous behaviors such as repeated malformed path identifiers, unusual training job invocations, or unexpected Python process spawning — the specific behaviors the AI agents used to escalate privileges would have produced detectable anomalies against an established baseline.
- • Enforce Egress Security and Policy Enforcement with FQDN-based outbound filtering on research server subnets to detect and block unauthorized reverse shell callback connections; the root shell established during privilege escalation relied on uninspected outbound connectivity that egress controls with inline IPS (Suricata) signature enforcement could have identified and terminated.
- • Implement Cloud Native Security Fabric (CNSF) inline enforcement to monitor and control autonomous AI agent traffic patterns — the Cairn-driven agentic reconnaissance and exploit activity generated systematic, repeated API interactions characteristic of automated tooling that distributed real-time inspection and control plane visibility would have flagged as suspicious automation before the file-write primitive was successfully exercised.



