Skip to main content
Vulnerability Database/CVE-2026-100846

CVE-2026-100846: MONAI Deserialization RCE Vulnerability

CVE-2026-100846 is a deserialization vulnerability in MONAI that enables remote code execution through malicious pickle files. This article covers the technical details, affected versions, and mitigation strategies.

Published:

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-2026-100846 Overview

CVE-2026-100846 is an insecure deserialization vulnerability [CWE-502] in Medical Open Network for AI (MONAI) versions prior to 1.5.2. The flaw resides in the algo_from_pickle function within monai/auto3dseg/utils.py. The function reads a .pkl file and passes its contents directly to pickle.loads without validating the data source or content. An attacker who supplies a crafted pickle file to an application calling algo_from_pickle can trigger execution of a Python object defining __reduce__, resulting in arbitrary code execution in the context of the host application.

Critical Impact

Successful exploitation grants arbitrary code execution within the MONAI application process, which commonly runs with access to medical imaging data and GPU-backed training infrastructure.

Affected Products

  • Project-MONAI MONAI versions prior to 1.5.2
  • Applications and pipelines invoking monai.auto3dseg.utils.algo_from_pickle
  • Auto3DSeg workflows ingesting externally sourced .pkl algorithm artifacts

Discovery Timeline

  • 2026-09-27 - CVE-2026-100846 published to the National Vulnerability Database (NVD)
  • 2026-09-30 - Last updated in NVD database

Technical Details for CVE-2026-100846

Vulnerability Analysis

MONAI is an open-source PyTorch-based framework for deep learning in healthcare imaging. The Auto3DSeg module automates 3D medical image segmentation pipelines and serializes algorithm configurations to pickle files on disk. The algo_from_pickle helper is responsible for reloading those saved algorithm objects.

The function opens the target .pkl file and invokes pickle.loads on the raw bytes with no provenance check, no integrity verification, and no restricted unpickler. Python's pickle module executes any callable referenced by an object's __reduce__ method during deserialization. An attacker who controls the contents of the loaded file can therefore run arbitrary Python code, spawn subprocesses, read datasets, or move laterally from the training host.

Root Cause

The root cause is trust placed in attacker-controllable serialized data. MONAI treats pickle files as internal artifacts, but algo_from_pickle does not enforce that assumption. There is no HMAC signature, no allow-list of modules via a custom Unpickler.find_class, and no sandbox around the deserialization call. Any workflow that accepts externally generated algorithm bundles inherits the risk.

Attack Vector

Exploitation requires an application built on MONAI to invoke algo_from_pickle on a file an attacker can influence. Realistic delivery paths include shared model zoos, collaborative research repositories, cloud object storage buckets, and user-uploaded Auto3DSeg algorithm archives. Once the malicious pickle is loaded, the embedded __reduce__ payload executes with the privileges of the MONAI process. See the GitHub Security Advisory GHSA-89gg-p5r5-q6r4 and the VulnCheck advisory for additional context.

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 Attempts

Detection Methods for CVE-2026-100846

Indicators of Compromise

  • Unexpected child processes (for example sh, bash, python, curl, wget) spawned by Python interpreters running MONAI workloads.
  • Outbound network connections originating from training or inference hosts shortly after loading an Auto3DSeg .pkl artifact.
  • .pkl files written to MONAI working directories from untrusted ingestion paths, user uploads, or external storage mounts.

Detection Strategies

  • Instrument Python applications to log every call to pickle.load and pickle.loads, including the source file path and process ancestry.
  • Hunt for process-tree anomalies where a MONAI/Auto3DSeg process is the parent of shell, interpreter, or reconnaissance utilities.
  • Scan artifact repositories and object storage for .pkl files containing suspicious opcodes such as R (REDUCE), c (GLOBAL), or i (INST) referencing os, subprocess, or builtins.

Monitoring Recommendations

  • Monitor file-system telemetry for new .pkl files landing in Auto3DSeg working directories from non-pipeline writers.
  • Track egress from GPU training nodes to unapproved destinations and alert on first-seen domains or IPs.
  • Correlate authentication and object-storage access logs with pipeline execution windows to spot untrusted artifact pulls.

How to Mitigate CVE-2026-100846

Immediate Actions Required

  • Upgrade MONAI to version 1.5.2 or later on every training, inference, and research host.
  • Inventory applications that call algo_from_pickle and restrict them to artifacts produced by trusted internal pipelines.
  • Quarantine any .pkl files received from external collaborators or public repositories pending integrity review.

Patch Information

Project-MONAI addressed the issue in release 1.5.2. Refer to the GitHub Security Advisory GHSA-89gg-p5r5-q6r4 for the fix details and upgrade guidance.

Workarounds

  • Do not invoke algo_from_pickle on files sourced from outside the trust boundary of the deploying organization.
  • Store Auto3DSeg artifacts in access-controlled locations and verify integrity with cryptographic signatures (for example HMAC or Sigstore) before loading.
  • Run MONAI workloads in sandboxed containers with least-privilege service accounts, read-only mounts for untrusted inputs, and egress filtering.
bash
# Upgrade MONAI to the patched release
pip install --upgrade "monai>=1.5.2"

# Verify 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

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.