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-2026-100843 Overview
CVE-2026-100843 is a remote code execution vulnerability in Medical Open Network for AI (MONAI) versions before 1.6.0. The flaw resides in the algo_from_pickle() function within monai/auto3dseg/utils.py, which calls Python's pickle.loads() on attacker-controlled data without validation. Deserializing a malicious pickle file causes arbitrary system commands to execute in the context of the MONAI process. The weakness is classified under CWE-502: Deserialization of Untrusted Data.
Critical Impact
Successful exploitation grants arbitrary code execution on the host running MONAI, placing medical imaging pipelines, model artifacts, and training datasets at risk of compromise.
Affected Products
- Project-MONAI MONAI versions prior to 1.6.0
- monai/auto3dseg/utils.py module containing algo_from_pickle()
- Downstream Auto3DSeg workflows that load serialized algorithm objects
Discovery Timeline
- 2026-09-27 - CVE-2026-100843 published to the National Vulnerability Database
- 2026-09-30 - Last updated in NVD database
Technical Details for CVE-2026-100843
Vulnerability Analysis
MONAI is a PyTorch-based framework for medical imaging research and deployment. The Auto3DSeg module automates 3D segmentation pipelines and persists algorithm state to disk using Python pickle files. The algo_from_pickle() helper reconstructs these algorithm objects by invoking pickle.loads() directly on file contents.
Python's pickle format is a stack-based virtual machine that permits arbitrary callable invocation during deserialization through opcodes such as REDUCE and GLOBAL. An attacker who places a crafted pickle file where MONAI will load it can execute arbitrary code the moment deserialization begins.
Exploitation requires local attack access and user interaction, typically a data scientist loading a shared or downloaded Auto3DSeg artifact. In collaborative research environments where models and intermediate artifacts move between workstations, clusters, and public repositories, this trust boundary is routinely crossed.
Root Cause
The root cause is the use of pickle.loads() on untrusted input without an allow-list of safe classes or a signature check. Python documentation explicitly warns against unpickling data from untrusted sources, yet the Auto3DSeg loader applied no mitigations such as pickle.Unpickler.find_class restrictions, HMAC verification, or migration to a safer format like safetensors or JSON.
Attack Vector
An attacker crafts a pickle payload whose __reduce__ method returns a callable such as os.system with a command string. The file is distributed through a model-sharing site, a shared network drive, a cloned Git repository, or a supply-chain package. When a MONAI user invokes an Auto3DSeg workflow that calls algo_from_pickle() against the malicious artifact, the embedded command runs with the privileges of the Python process.
See the Project-MONAI GitHub Security Advisory GHSA-qxq5-qhx6-94qw and the VulnCheck advisory for MONAI RCE for 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-2026-100843
Indicators of Compromise
- Unexpected child processes spawned by Python interpreters running MONAI, particularly shells (sh, bash, cmd.exe, powershell.exe) with no corresponding user activity.
- Pickle files (.pkl, .pickle) in Auto3DSeg working directories originating from untrusted sources or recently modified without a corresponding training run.
- Outbound network connections initiated by the MONAI Python process to unfamiliar hosts shortly after loading an algorithm artifact.
Detection Strategies
- Inventory installations of the monai package across research and production environments and flag versions below 1.6.0.
- Instrument runtime hooks to log every invocation of pickle.loads() within MONAI code paths and alert on calls referencing non-whitelisted files.
- Correlate process-ancestry telemetry to detect Python processes executing operating-system commands immediately after reading a .pkl file.
Monitoring Recommendations
- Monitor file-system events for writes of .pkl files into Auto3DSeg directories by non-pipeline users or services.
- Alert on MONAI processes performing egress connections, filesystem traversal outside dataset paths, or credential-access behavior.
- Review audit logs for shared model repositories to identify who uploaded or modified serialized algorithm objects.
How to Mitigate CVE-2026-100843
Immediate Actions Required
- Upgrade MONAI to version 1.6.0 or later across all workstations, training clusters, and container images.
- Audit existing Auto3DSeg artifacts and quarantine any pickle files whose provenance cannot be verified.
- Restrict write access to directories consumed by algo_from_pickle() to trusted pipeline identities only.
Patch Information
Project-MONAI addressed the issue in MONAI 1.6.0. Review the GitHub Security Advisory GHSA-qxq5-qhx6-94qw for upgrade guidance and the VulnCheck advisory for additional context.
Workarounds
- Avoid loading Auto3DSeg pickle files from any source outside your organization's trust boundary until the upgrade is complete.
- Run MONAI workloads inside least-privilege containers or sandboxes with restricted network egress and no access to secrets.
- Where feasible, re-serialize trusted algorithm state in a safer format such as safetensors or JSON and remove legacy pickle artifacts.
# Upgrade MONAI to a fixed release
pip install --upgrade "monai>=1.6.0"
# Verify the installed version
python -c "import monai; print(monai.__version__)"
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

