Skip to main content
Vulnerability Database/CVE-2026-108945

CVE-2026-108945: iMoonLab yolov13 RCE Vulnerability

CVE-2026-108945 is a code injection flaw in iMoonLab yolov13 affecting the AutoBackend Model Loading component. Attackers can exploit argument manipulation to execute arbitrary code. This article covers technical details, impact analysis, 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-2026-108945 Overview

CVE-2026-108945 is a code injection vulnerability [CWE-74] in the iMoonLab YOLOv13 machine learning project. The flaw resides in the eval function call within ultralytics/nn/autobackend.py, part of the AutoBackend Model Loading component. An attacker with local access can manipulate the imgsz, names, or kpt_shape arguments to inject arbitrary code during model loading. The vulnerability affects iMoonLab YOLOv13 up to commit 73289949533efac82bb5f72ec19b746618656bd2. The maintainers were notified via a GitHub issue report but have not yet responded.

Critical Impact

Local attackers can execute arbitrary Python code during model loading by supplying a crafted YOLOv13 model with malicious metadata fields.

Affected Products

  • iMoonLab YOLOv13 up to commit 73289949533efac82bb5f72ec19b746618656bd2
  • Component: ultralytics/nn/autobackend.py AutoBackend Model Loading
  • Downstream projects embedding the vulnerable YOLOv13 code path

Discovery Timeline

  • 2026-10-13 - CVE-2026-108945 published to NVD
  • 2026-10-15 - Last updated in NVD database

Technical Details for CVE-2026-108945

Vulnerability Analysis

The vulnerability is a code injection flaw [CWE-74] in the YOLOv13 AutoBackend model loader. During model deserialization, the eval function in ultralytics/nn/autobackend.py processes model metadata fields including imgsz, names, and kpt_shape. These fields are evaluated as Python expressions rather than parsed as structured data. An attacker who controls a model file loaded by the application can embed arbitrary Python expressions into any of these fields. When the application loads the model, the embedded expressions execute in the Python interpreter running the YOLOv13 process. Execution occurs with the privileges of the user running the loader.

Root Cause

The root cause is the use of Python's built-in eval on attacker-controllable model metadata. eval interprets a string as a Python expression and executes it, so any input reaching that call becomes code. Safe alternatives such as ast.literal_eval or explicit type parsing were not used for these fields. The AutoBackend loader also lacks input validation on imgsz, names, and kpt_shape before passing them to eval.

Attack Vector

The attack requires local access and low privileges. An attacker crafts a YOLOv13-compatible model file in which the imgsz, names, or kpt_shape metadata contains a Python expression such as a call to os.system or __import__. When a victim loads the crafted model through AutoBackend, the expression is passed to eval and the injected code runs. No user interaction beyond loading the model is required. See the GitHub Issue #81 Discussion for additional context on the reported behavior.

// No verified exploit code is available. Refer to the vendor issue tracker for proof-of-concept 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 Attempts

Detection Methods for CVE-2026-108945

Indicators of Compromise

  • Unexpected child processes spawned by Python interpreters running YOLOv13 or Ultralytics-derived code
  • Model files containing non-numeric, expression-like strings in imgsz, names, or kpt_shape fields
  • Outbound network connections initiated shortly after a model load operation
  • New files written to the working directory or temp paths by the model loader process

Detection Strategies

  • Perform static analysis of model files prior to loading and reject fields that do not match expected primitive types
  • Instrument Python processes to log calls to eval, exec, os.system, and subprocess originating from autobackend.py
  • Hash and whitelist trusted model artifacts, alerting when an unapproved model is loaded
  • Monitor for process lineage anomalies where Python spawns shells or system utilities

Monitoring Recommendations

  • Enable endpoint telemetry on hosts and workstations used by ML engineers for model inference
  • Collect process creation and command-line arguments for Python interpreters loading YOLOv13 models
  • Forward Python audit hook events (PEP 578) for exec, compile, and eval into a centralized log store for correlation

How to Mitigate CVE-2026-108945

Immediate Actions Required

  • Restrict model loading to files sourced from trusted internal repositories only
  • Run YOLOv13 inference under a dedicated low-privilege account with no network egress where feasible
  • Inspect existing model inventories for metadata fields that contain callable syntax rather than literals
  • Isolate ML workstations and inference hosts from sensitive internal services

Patch Information

No vendor patch has been released. According to the VulDB CVE-2026-108945 entry, the project was informed of the issue through a GitHub report but has not responded. Monitor the GitHub YOLOv13 Repository and GitHub Issue #81 Discussion for updates. Downstream users should consider forking and replacing eval with ast.literal_eval in ultralytics/nn/autobackend.py.

Workarounds

  • Replace the eval call in ultralytics/nn/autobackend.py with ast.literal_eval to restrict parsing to Python literals
  • Validate imgsz, names, and kpt_shape against strict schemas before they reach the loader
  • Sandbox model loading using containerization or seccomp profiles that block process execution and network syscalls
  • Require cryptographic signature verification for model files before loading
bash
# Example: run YOLOv13 inference inside a restricted container with no network and read-only model mounts
docker run --rm --network=none --read-only \
  --cap-drop=ALL --security-opt=no-new-privileges \
  -v /opt/models:/models:ro \
  yolov13-runtime:restricted python infer.py --weights /models/trusted.pt

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.