Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-100840 Overview
CVE-2026-100840 is a remote code execution vulnerability affecting Medical Open Network for Artificial Intelligence (MONAI) through version 1.6.0. The flaw resides in the bundle configuration engine, which resolves _target_ values to arbitrary importable callables without an allow list. The engine also passes $ expressions directly to Python eval(). Attackers can publish a malicious bundle containing crafted configuration that executes arbitrary code when a victim loads it with monai.bundle.load() or monai.bundle.run(). The weakness maps to CWE-95: Improper Neutralization of Directives in Dynamically Evaluated Code.
Critical Impact
Loading an untrusted MONAI bundle results in arbitrary code execution in the user's Python process, compromising confidentiality, integrity, and availability of the host.
Affected Products
- Project-MONAI MONAI versions through 1.6.0
- monai.bundle.load() consumers
- monai.bundle.run() consumers
Discovery Timeline
- 2026-09-27 - CVE-2026-100840 published to NVD
- 2026-09-30 - Last updated in NVD database
Technical Details for CVE-2026-100840
Vulnerability Analysis
MONAI is a widely used PyTorch-based framework for medical imaging AI. Its bundle format packages model weights, metadata, and configuration into a distributable artifact. The bundle configuration engine supports two features that create the code execution path: _target_ keys that reference Python callables, and $ prefixed expressions evaluated at runtime.
When a user calls monai.bundle.load() or monai.bundle.run(), the engine resolves _target_ values by importing any dotted path the configuration specifies. There is no allow list restricting which modules or callables may be instantiated. An attacker can therefore target callables such as os.system, subprocess.Popen, or builtins.exec and supply arbitrary arguments through the bundle configuration.
The $ expression handler compounds the issue by routing expression strings through Python's eval() built-in. This evaluates attacker-controlled Python source in the context of the loading process, bypassing any argument-level restrictions. Combined, these two primitives give a bundle author full code execution on any machine that opens the bundle.
Root Cause
The root cause is unsafe dynamic code evaluation in trusted deserialization paths. The configuration parser treats bundle content as trusted input and provides general-purpose Python execution primitives without sandboxing, allow-listing, or signature verification.
Attack Vector
Exploitation requires a victim to load an attacker-supplied bundle. Distribution channels include model hubs, shared research artifacts, container images bundling pretrained weights, and internal MLOps pipelines that ingest third-party checkpoints. Once the victim invokes monai.bundle.load() or monai.bundle.run(), the malicious _target_ resolution or $ expression executes immediately in the Python process.
See the VulnCheck Advisory for MONAI RCE and the GitHub Security Advisory GHSA-873f-pvrv-4x83 for additional technical details.
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-100840
Indicators of Compromise
- Bundle configuration files (*.yaml, *.json) containing _target_ values resolving to os.system, subprocess.*, builtins.exec, builtins.eval, or importlib primitives.
- Configuration values prefixed with $ that contain Python expressions invoking network, filesystem, or process APIs.
- Unexpected child processes spawned by Python interpreters that previously imported monai.bundle.
- Outbound network connections initiated from research workstations or training nodes shortly after bundle load events.
Detection Strategies
- Perform static analysis of bundle archives before loading, flagging any _target_ value outside an approved MONAI or PyTorch namespace.
- Scan configuration files for $ prefixed expressions and route them for manual review.
- Correlate monai.bundle.load or monai.bundle.run invocations with subsequent process creation events on the same host.
Monitoring Recommendations
- Monitor data science hosts for Python processes spawning shells, package managers, or network utilities.
- Capture command-line arguments and parent process IDs for all child processes of Python interpreters used in ML workflows.
- Audit model registry and artifact repository access logs for pulls from untrusted sources.
How to Mitigate CVE-2026-100840
Immediate Actions Required
- Treat all MONAI bundles sourced outside your organization as untrusted code and quarantine them pending review.
- Upgrade MONAI to a fixed release as published in the GitHub Security Advisory GHSA-873f-pvrv-4x83.
- Inventory existing bundles in model registries and scan their configuration for _target_ and $ abuse patterns.
- Run bundle loading operations under least-privilege service accounts with no outbound internet access by default.
Patch Information
Project-MONAI has published fix guidance in GHSA-873f-pvrv-4x83. Users of MONAI 1.6.0 and earlier should upgrade to the patched release referenced in the advisory and rebuild any downstream container images or Python environments that pin the vulnerable version.
Workarounds
- Load bundles only from sources you control or cryptographically verify.
- Execute bundle load and run operations inside sandboxed containers with seccomp and network egress restrictions.
- Pre-process bundle configuration to strip or reject $ expressions and non-allow-listed _target_ values.
- Require code review and signed provenance for every bundle promoted into shared or production registries.
# Configuration example: scan a bundle for dangerous constructs before loading
grep -REn '"_target_"\s*:\s*"(os|subprocess|builtins|importlib|sys)\.' ./bundle/
grep -REn '"\$' ./bundle/
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

