Executive Summary
Between September 8–10, 2026, a Chinese-nexus threat actor weaponized Cairn, a legitimate autonomous AI pentesting platform, to discover and exploit a path traversal zero-day vulnerability (CVE tracked under GHSA-qc22-rmgc-vg94) in MONAI Label, an open-source medical imaging application. Within 11 minutes of mapping the application's API, Cairn's AI agents wrote a Python reverse shell to a system startup file, achieving root-level server takeover at targeted East Asian universities. Amazon Threat Intelligence uncovered the campaign after the attackers inadvertently left their Cairn instance publicly exposed without authentication, revealing the full scope of their autonomous hacking operations.
This incident marks a documented inflection point: AI agents autonomously discovered and weaponized a previously unknown vulnerability faster than any human red-team could respond. As autonomous offensive AI tools proliferate beyond research labs and into active threat-actor toolkits, defenders face a shrinking window between vulnerability existence and exploitation—making AI-aware, automated network controls no longer optional.
Why This Matters Now
Autonomous AI exploitation tools like Cairn compress the zero-day discovery-to-exploitation timeline from weeks to minutes, rendering traditional patch-and-scan cycles obsolete. With nation-state actors now delegating offensive operations to AI agents targeting critical-sector infrastructure—including healthcare and academia—organizations must deploy always-on, automated network segmentation and egress controls capable of neutralizing threats before human analysts can even triage an alert.
Attack Path Analysis
A Chinese-nexus threat actor deployed the autonomous AI agent platform Cairn to discover and exploit a path traversal zero-day in MONAI Label, a medical imaging application, gaining root shell access to a university server within 11 minutes of API reconnaissance. The agents subsequently pivoted via attacker-supplied proxy into a second university network running the same vulnerable software, demonstrating AI-accelerated lateral movement and multi-target exploitation that outpaced manual defensive response.
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
Cairn AI agents were directed to target internet-exposed MONAI Label instances at East Asian educational institutions. The agents autonomously mapped the application's API surface and discovered a previously unknown path traversal zero-day vulnerability (later assigned GHSA-qc22-rmgc-vg94), enabling arbitrary file writes to the server. Within 11 minutes of beginning API enumeration, the agents transitioned from reconnaissance to active exploitation of the zero-day flaw.
Related CVEs
CVE-2025-49595
CVSS 4.9A path traversal vulnerability in MONAI Label allows unauthenticated remote attackers to write files to arbitrary locations on the server, enabling remote code execution including deployment of reverse shells with root-level access.
Affected Products:
Project MONAI MONAI Label – < 0.5.3
Exploit Status:
exploited in the wildReferences:
https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94https://nvd.nist.gov/vuln/detail/CVE-2025-49595https://www.bleepingcomputer.com/news/security/hackers-use-autonomous-ai-tool-cairn-to-find-zero-day-take-over-server/https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/
MITRE ATT&CK® Techniques
Exploit Public-Facing Application
Exploitation for Client Execution
Command and Scripting Interpreter: Python
Server Software Component: Web Shell
Abuse Elevation Control Mechanism
File and Directory Discovery
Brute Force: Password Guessing
Application Layer Protocol: Web Protocols
Potential Compliance Exposure
Mapping incident impact across multiple compliance frameworks.
NIST SP 800-53 Rev 5 – Flaw Remediation and Vulnerability Scanning
Control ID: SI-2 / RA-5
NIS2 Directive (EU) 2022/2555 – Vulnerability Handling and Disclosure
Control ID: Article 21(2)(e)
CISA Zero Trust Maturity Model 2.0 – Application-Level Access Control and Automated Threat Response
Control ID: Application Pillar – Visibility & Analytics / Automation & Orchestration
DORA (EU) 2022/2554 – Threat-Led Penetration Testing and Vulnerability Management
Control ID: Article 10 – ICT Vulnerability Management
HIPAA Security Rule (45 CFR Part 164) – Protection of Electronic Protected Health Information
Control ID: 164.306(a)(1) / 164.312(a)(2)(iv)
ISO/IEC 27001:2022 – Technical Vulnerability and Patch Management
Control ID: Annex A 8.8 – Management of Technical Vulnerabilities
Sector Implications
Industry-specific impact of the vulnerabilities, including operational, regulatory, and cloud security risks.
Health Care / Life Sciences
AI-assisted zero-day exploitation of MONAI Label medical imaging software enabled root server takeover, exposing HIPAA.164.312(e)(1) and PCI compliance gaps in unprotected research environments.
Higher Education/Acadamia
Chinese-nexus Cairn AI agents directly targeted East Asian university GPU servers, exploiting lateral movement vectors and triggering zero trust segmentation failures across academic multicloud infrastructure.
Research Industry
Autonomous AI exploitation of open-source medical imaging tools used by research institutions highlights critical east-west traffic security and egress policy enforcement gaps in unmonitored lab environments.
Computer/Network Security
Cairn's autonomous 11-minute zero-day discovery cycle outpaces manual defenses, demanding cloud-native security fabric deployment, inline IPS, and anomaly detection to counter AI-driven threat actors.
Sources
- Hackers use autonomous AI tool Cairn to find zero-day, take over serverhttps://www.bleepingcomputer.com/news/security/hackers-use-autonomous-ai-tool-cairn-to-find-zero-day-take-over-server/Verified
- Chinese-Nexus Actors: Cairn AI Agents Found a Zero-Day on Their Own - AWS Security Bloghttps://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 - Path Traversal Zero-Dayhttps://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94Verified
- Malicious AI agents steal 600K credit cards, infect 100+ sites with skimmershttps://www.bleepingcomputer.com/news/security/malicious-ai-agents-steal-600k-credit-cards-infect-100-plus-sites-with-skimmers/Verified
- Google warns that fully-automated AI attacks are coming soonhttps://www.bleepingcomputer.com/news/security/google-warns-that-fully-automated-ai-attacks-are-coming-soon/Verified
- DIVD says Zammad zero-days enabled AI-driven network breachhttps://www.bleepingcomputer.com/news/security/divd-says-zammad-zero-days-enabled-ai-driven-network-breach/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 Cairn-driven attack chain depended on unrestricted internet reachability to exposed MONAI Label instances and uncontrolled east-west paths between institution networks. By enforcing identity-aware segmentation and controlled egress, CNSF would likely reduce the attacker's ability to pivot freely across network boundaries and limit the blast radius of the initial compromise.
Control: Cloud Native Security Fabric (CNSF)
Mitigation: Internet reachability to the exposed MONAI Label API would likely be reduced, limiting the AI agents' ability to freely enumerate and probe the application's attack surface from arbitrary external origins.
Control: Zero Trust Segmentation
Mitigation: Even with file write achieved, Zero Trust segmentation would likely constrain the scope of what the now-compromised workload could reach, reducing the effective blast radius of the root shell beyond the initially targeted server.
Control: East-West Traffic Security
Mitigation: East-west traffic enforcement would likely constrain the agents' ability to traverse freely between institution networks and reach additional MONAI Label workloads via the attacker-supplied proxy path.
Control: Multicloud Visibility & Control
Mitigation: Centralized multicloud visibility and control would likely surface anomalous outbound connection attempts from the MONAI Label workload, and policy enforcement could constrain the reverse shell's ability to maintain a persistent outbound channel to attacker infrastructure.
Control: Egress Security & Policy Enforcement
Mitigation: Egress policy enforcement would likely constrain the available outbound paths from the compromised server, reducing the feasibility of bulk data transfer to attacker-controlled destinations even with root access established.
With prior segmentation and egress controls in place, the residual impact would likely be reduced to a more contained server-level compromise rather than a multi-institution takeover, limiting the attacker's ability to operationalize root access across connected environments.
Impact at a Glance
Affected Business Functions
- Medical Imaging Research and Annotation Pipelines
- Academic GPU Computing Infrastructure
- Research Data Repositories and Patient-Adjacent Imaging Datasets
- University Network Lateral Movement and Credential Stores
Estimated downtime: 7 days
Estimated loss: N/A
Attackers achieved root-level shell access on servers hosting MONAI Label, an open-source medical imaging AI annotation platform used in academic and research settings. Data at risk includes de-identified or potentially re-identifiable medical imaging datasets (e.g., MRI, CT, pathology scans), research intellectual property associated with AI model training pipelines, server credentials and SSH keys stored on compromised hosts, and potentially other connected university network resources accessible via lateral movement following the initial server takeover. The operators subsequently pivoted to a second university network running the same vulnerable software, amplifying exposure scope across multiple institutions.
Recommended Actions
Key Takeaways & Next Steps
- • Deploy Cloud Native Security Fabric (CNSF) with inline enforcement and real-time API traffic inspection to detect anomalous autonomous agent behaviors — such as rapid sequential API enumeration and repeated malformed path traversal requests — that characterize AI-driven zero-day discovery before exploitation completes.
- • Enforce Zero Trust Segmentation with least-privilege identity-based policies between application workloads to contain blast radius; prevent compromised MONAI Label instances from communicating laterally to other university network segments or GPU servers without explicit policy authorization.
- • Implement Egress Security and Policy Enforcement with FQDN filtering and outbound traffic controls to block unauthorized reverse shell callback connections from application servers to attacker-controlled infrastructure, interrupting Command & Control before persistent root access is established.
- • Apply East-West Traffic Security controls with workload-to-workload flow inspection and microsegmentation to prevent attacker-supplied proxy pivots from enabling AI agents to traverse from a compromised first institution into the second university network running the same vulnerable software.
- • Activate Multicloud Visibility and anomaly detection with traffic observability baselines to alert on suspicious automation patterns — including high-frequency API calls, rapid file write operations, and unexpected outbound connection initiation from medical imaging application processes — enabling faster-than-manual automated defensive response against AI-accelerated attack chains.



