CVE-2026-119578
MONAI Auto3DSeg imports code from untrusted algorithm work directories
Project MONAI · MONAI Auto3DSeg algo_from_json
First documented disclosure · Review checkpoint
Summary
Loading an untrusted algorithm work directory can execute Python during construction.
Severity and exploitation evidence
High · 8.5/10 · CVSS 4.0. Severity source.
Exploitation not established: The original advisory demonstrates controlled code execution from a crafted work directory. No reviewed source establishes malicious exploitation or confirmed victims.
Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.
How the weakness works
algo_object.json controls a target class and template paths added to Python's import search path; the loader imports and instantiates the target.
Conditions required
A user loads an attacker-supplied work directory with AutoRunner or import_bundle_algo_history. The demonstrated JSON path does not require MONAI_ALLOW_PICKLE.
Potential impact
Code execution with the workload user's filesystem and credential access.
Affected versions and configurations
CNA: >=1.6.0 and <1.6.1. The original GHSA instead lists <=1.6.0rc1; its range metadata differs.
Fix and mitigation status
1.6.1 is vendor-listed and restricts targets to Algo subclasses, but the patch explicitly retains unsafe module-import risk for untrusted template directories.
Practical next steps
Use a supported patched release and load only trusted algorithm bundles in an isolated environment. Do not treat the subclass restriction as a safe loader for attacker-controlled modules. Track the later literal-state fix separately.
Operating lesson
Type-checking a class after importing its module does not make the import safe.
Evidence and limits
September 3 PR 9085 disclosed the exact sink and added a warning. September 11 PR 9115, merged September 17, narrows targets but explicitly says importing still executes top-level code. The October 15 GHSA and October 30 CNA list 1.6.1 as fixed; that does not erase this limitation. October 29 PR 9233 separately makes saved-state strings literal; no fixed release for that later change is inferred. CVSS 4.0 8.5 is the CNA assessment, versus 7.8 CVSS 3.1.
Sources
- Public warning-only security PR
Primary source ·
- Target restriction and explicit residual risk
Primary source ·
- Maintainer original advisory
Primary source ·
- Later literal-state hardening
Primary source ·
- Exact published CVE authority record
Primary source ·