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-2026-111865 Overview
CVE-2026-111865 is a code injection vulnerability in the THU-MIG yoloe machine learning project. The flaw resides in the eval function within ultralytics/nn/autobackend.py, part of the AutoBackend Model Loading component. Attacker-controlled values passed through the imgsz, names, or kpt_shape arguments reach a dynamic code evaluation path, enabling improper neutralization of directives in dynamically evaluated code [CWE-94]. Exploitation requires local access and low privileges. The project does not use versioning, so no fixed release is identified; the affected code is tracked through commit 40cd606cabdbe2b566d6f14a6b162c89206e9a1b. Public exploit information has been released.
Critical Impact
Local attackers with low privileges can execute arbitrary Python code when a crafted model is loaded through the AutoBackend component.
Affected Products
- THU-MIG yoloe repository up to commit 40cd606cabdbe2b566d6f14a6b162c89206e9a1b
- Component: AutoBackend Model Loading (ultralytics/nn/autobackend.py)
- Vulnerable function: eval processing imgsz, names, and kpt_shape arguments
Discovery Timeline
- 2026-10-17 - CVE-2026-111865 published to the National Vulnerability Database (NVD)
- 2026-10-19 - Last updated in the NVD database
Technical Details for CVE-2026-111865
Vulnerability Analysis
The vulnerability is a Code Injection flaw [CWE-94] in the AutoBackend model loading routine of THU-MIG yoloe. When a model is loaded, metadata fields including imgsz (image size), names (class names), and kpt_shape (keypoint shape) are parsed from the model file and passed through Python's eval built-in. Because eval compiles and executes arbitrary expressions, any attacker who controls the contents of a model file can embed Python code that runs under the identity of the user loading the model. The result is local arbitrary code execution within the ML inference environment.
The project does not use semantic versioning, so administrators cannot pin to a safe release. Reviewing the referenced GitHub Project Repository and GitHub Issue #148 Discussion is required to identify the current state of the code.
Root Cause
The root cause is the use of eval on untrusted serialized model metadata. Values that should be parsed with a safe literal evaluator, such as ast.literal_eval, are instead evaluated as live Python expressions. Any string embedded in imgsz, names, or kpt_shape is treated as executable code.
Attack Vector
Exploitation requires a local approach. An attacker crafts a malicious model file in which the serialized imgsz, names, or kpt_shape fields contain Python expressions rather than plain data. When a user, data scientist, or automation pipeline loads the model with yoloe AutoBackend, the embedded expression is executed. Common delivery paths include shared model registries, public model hubs, and model files distributed in notebooks or training datasets.
No verified exploit code is reproduced here. Technical details are tracked at VulDB CVE-2026-111865.
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-2026-111865
Indicators of Compromise
- Unexpected Python child processes (for example sh, bash, curl, python -c) spawned from a process loading a yoloe or ultralytics model
- Model files whose metadata fields contain non-literal strings such as function calls, imports, or __import__ references
- Outbound network connections initiated immediately after model load events on inference hosts or developer workstations
Detection Strategies
- Scan model artifacts for Python keywords (__import__, exec, eval, os.system, subprocess) inside imgsz, names, or kpt_shape fields prior to loading
- Alert on process lineage where a Python interpreter running autobackend.py spawns shell interpreters or network utilities
- Correlate file-write events in ultralytics/nn/autobackend.py with repository state to detect tampering in local checkouts
Monitoring Recommendations
- Enable command-line and process-tree logging on hosts that perform ML inference or model conversion
- Monitor model registry pull events and compare file hashes against trusted baselines
- Track egress traffic from GPU and inference nodes, which typically have narrow, predictable network profiles
How to Mitigate CVE-2026-111865
Immediate Actions Required
- Treat all third-party yoloe model files as untrusted until the upstream project responds to GitHub Issue #148 Discussion
- Restrict model loading to isolated, non-privileged accounts without access to production credentials or secrets
- Audit existing model repositories for files whose metadata contains executable Python expressions
Patch Information
No vendor patch is available at the time of publication. The maintainers were notified through an issue report but have not responded. Because the project does not use versioning, affected and unaffected releases cannot be enumerated. Track the GitHub Project Repository for remediation commits that replace eval with a safe parser such as ast.literal_eval.
Workarounds
- Load models only inside containers or sandboxes with no network egress and read-only filesystem access to sensitive paths
- Apply a local patch to ultralytics/nn/autobackend.py substituting ast.literal_eval for eval when parsing imgsz, names, and kpt_shape
- Enforce code review and signature verification for every model artifact before it is deployed to shared inference infrastructure
- Remove write access to model directories from accounts that do not require it
# Example sandboxed invocation using a non-privileged user and no network
sudo -u ml-sandbox \
firejail --net=none --read-only=/etc --private-tmp \
python -c "from ultralytics.nn.autobackend import AutoBackend; AutoBackend('model.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

