Autonomous Agents Weaponize MONAI Label Zero-Day to Seize Root Shells in 11 Minutes

Amazon Threat Intelligence discloses CVE-2026-133968, an unauthenticated CVSS 9.8 zero-day in medical imaging servers exploited by autonomous Cairn agents.

Published: 2026.11.22

Visit Official Site
Recon to Arbitrary Write
11 Minutes

Autonomous Cairn agents discovered API endpoints and wrote payload files

Vulnerability Severity
CVSS 9.8

Unauthenticated remote path traversal in MONAI Label 0.8.5 and earlier

Tagged Upstream Release
0 Patches

Source fix merged to master; no official distribution release published

Executive Takeaway ·Deploying medical AI frameworks with default zero-auth configs and privileged Docker sockets allows autonomous red-teaming agents to convert file traversa…

Autonomous Reconnaissance to Arbitrary Disk Writes: The Mechanized Anatomy of CVE-2026-133968

Autonomous software agents have transitioned from theoretical penetration testing concepts into active offensive operations. Amazon Threat Intelligence documented an intrusion where a suspected China-linked operator deployed autonomous agents via the open-source Cairn framework against academic research infrastructure. The target environment ran MONAI Label, an open-source server within the broader Project MONAI ecosystem used to annotate medical scans and train deep learning models.

The underlying vulnerability, indexed as CVE-2026-133968 with a CVSS 9.8 critical rating, is an unauthenticated directory traversal flaw in MONAI Label 0.8.5 and earlier. The server exposes image and label upload endpoints that accept user-controlled file identifiers. The application logic appends these identifiers directly to its internal storage root without verifying whether the resolved canonical path remains bounded within the target directory.

Because MONAI Label deploys with network authentication disabled by default, any client reaching the exposed port can trigger arbitrary file writes. The agents received high-level objectives from their operator—specifically to map target networks and capture interactive shells—without manual exploitation scripts. Operating against an internet-facing MONAI Label instance, the agents systematically enumerated REST endpoints and achieved out-of-bounds disk writes within 11 minutes of discovery.

Autonomous Exploitation Sequence of CVE-2026-133968

11-minute reconnaissance leading to root shell execution

01 · Step 01
API Enumeration

Autonomous Cairn agents scan exposed MONAI Label ports and identify unauthenticated REST endpoints

02 · Step 02
Path Traversal Write

Agents submit unsanitized file identifiers to escape storage directories and write outside designated folders

03 · Step 03
Execution Hijack

Malicious payload planted in Python auto-import paths triggered during process startup

04 · Step 04
Task Invocation

Agents trigger native model training API, spawning a new Python worker running under the server process

05 · Step 05
Root Reverse Shell

Worker process executes payload in privileged container context, returning an external reverse shell

Empirical Attack Profile & Blast Radius Matrix

Offensive AI agents lower the cost and latency of target reconnaissance, parameter testing, and lateral staging. Analysis of the operator’s exposed Cairn control panel revealed an iterative runtime loop: agents parsed application documentation, structured custom HTTP payloads, evaluated response codes, and adjusted path traversal syntax without human intervention.

The critical severity stems from the confluence of absent transport authentication and permissive container defaults. In academic and enterprise research labs, deep learning training rigs frequently run as containerized root services to facilitate direct access to host CUDA drivers, NVLink buses, and physical GPUs. When an unauthenticated path traversal intersects with root container permissions, remote code execution inevitably follows.

Operational VectorStandard Exploitation (Human-Led)Autonomous Agent Exploitation (Cairn)Enterprise Blast Radius
Reconnaissance to Write4–24 hours (manual fuzzing)11 minutes (automated API mapping)Rapid payload placement before SOC detection
Authentication BarrierExploits stolen credentials/tokensLeverages default zero-auth endpointsTrivial penetration on exposed default ports
Vulnerability ClassKnown CVE weaponizationZero-day discovery (CVE-2026-133968)Circumvents signature-based perimeter WAFs
Execution ContextService account isolationService account inherits container rootFull host and storage sub-system compromise
Upstream AvailabilityOfficial vendor patch deployedFix merged in git; no release tagSystems running PyPI 0.8.5 remain vulnerable
Average Incident TCO$180,000 (standard breach triage)$420,000+ (PHI/model audit & rebuild)Extensive forensic invalidation of dataset integrity

The table above details the operational shift. Conventional red-team engagements and targeted threat campaigns exhibit dwell times governed by manual tool execution and human analysis. Autonomous frameworks compress reconnaissance, fuzzed input generation, and exploitation into single-digit minute windows, invalidating static perimeter defenses that rely on human-speed alerts.

The Persistence Pipeline: From Directory Traversal to Root Code Execution via Python Hooks

A directory traversal vulnerability restricted to arbitrary file writing does not inherently yield code execution. The attacker must deposit data into a file path that is subsequently executed by an existing system process. In this intrusion, the agents resolved this constraint by exploiting Python’s runtime initialization sequence.

The agents directed the traversal payload to deposit a malicious script into a directory evaluated by the Python runtime on boot (such as site-packages or auto-executed configuration hooks). Python environments automatically execute .pth files and startup configurations when launching new interpreters. Once the payload sat on disk, the agents invoked a legitimate MONAI Label feature: triggering a model fine-tuning job via the REST API.

Exploitation Velocity and Containment Window Comparison

Latency metrics comparing human-driven attacks against autonomous agent execution

Human-Led Red Team Autonomous Cairn Agent
API Mapping to Payload Injection 29x faster execution
Human-Led Red Team 320 min
Autonomous Cairn Agent 11 min
Reverse Shell Establishment 30x shorter dwell window
Human-Led Red Team 1,440 min
Autonomous Cairn Agent 48 min

The fine-tuning call spawned a secondary Python worker process. That new process loaded the planted startup file, executed the embedded payload, and initiated an outbound network connection back to the attacker’s listener. Because the primary MONAI Label container ran without user namespace mapping or non-root privileges, the spawned process executed with full root authority.

This multi-stage vector bypassed conventional static analysis rules. The incoming network traffic appeared as typical REST traffic: an unauthenticated image upload followed by a standard dataset training command. No direct shell injection payloads traversed the edge proxy; the operating system executed the compromise natively from within its legitimate operational workflow.

Defense Playbook and Risk Remediation Matrix for Medical AI Infrastructure

Project maintainers merged a source-level fix on October 19. The commit updates path resolution functions to strictly enforce that target paths reside within designated storage roots and rejects dot-prefixed filenames or empty inputs. However, upstream maintainers have not yet tagged an official release beyond version 0.8.5, which dates back to 2024. Engineering teams deploying MONAI Label via pip packages or public container images remain unpatched by default.

Immediate Mitigation Decision Framework for MONAI Label
Can your team build custom containers directly from the project master branch?
Yes (DevOps / Container Build Pipeline Available)
Rebuild from Git Commit

Pull the latest master commit containing the October 19 patch; recompile internal container images.

→ Production AI Pipelines
No (Relying on Off-the-Shelf PyPI / Docker Hub Images)
Implement Strict Network Hardening

Isolate MONAI Label behind a VPN/mTLS proxy, enforce non-root UID 10001, and mount file stores as read-only.

→ Academic Labs & Test Environments

Infrastructure operators running medical imaging and ML training platforms must enact immediate defensive postures across four operational vectors:

  • Egress and Ingress Network Segmentation: Remove all public IP bindings from MONAI Label instances. Enforce edge authentication using mutual TLS (mTLS) or enterprise Identity-Aware Proxies (IAP). Medical labeling interfaces should never accept public ingress without authenticated identity tokens.
  • Service Account Demotion: Strip container images of root privileges. Configure container runtimes to drop CAP_SYS_ADMIN capabilities, enforce runAsNonRoot: true, and assign static non-root user IDs (e.g., UID 10001). If a process cannot write to Python site-packages or host system directories, traversal write exploits cannot achieve command execution.
  • Filesystem Read-Only Enforcements: Mount the underlying Python environment, system binaries, and configuration directories as read-only volumes (readOnlyRootFilesystem: true). Only the precise scratch storage folder assigned to training images should retain write permissions.
  • Source-Level Patching: In the absence of an official PyPI version 0.8.6 release, update CI/CD pipelines to install MONAI Label directly from the project’s default branch on GitHub rather than pulling the vulnerable 0.8.5 package from public package registries.

Related Benchmark Reports

View All in Category →
Weekly Briefing

Weekly Tech & Business Data Briefing

Verified software analysis, practical gotchas, and essential supply-chain updates delivered weekly.

Unsubscribe with 1 click anytime. Zero spam.

* We may earn an affiliate commission from links in this report, at no extra cost to you and with zero impact on our benchmark data.