WeaknessASC-2026-0061Other weakness
Remote code in model repositories (trust_remote_code)
MediumMitigated
Loading some models runs Python code from the model repository, through trust_remote_code or config auto_map entries, giving the repository's author code execution.
How to detect it
- trust_remote_code=True in loading code
- Model configs with auto_map pointing to custom code (vLLM CVE-2025-66448)
How to resolve it
- Keep trust_remote_code off; use models supported natively
- Pin and review any remote code you must run
- Update inference servers that executed it implicitly
How to investigate
- Find loads with remote code enabled and the code they ran
- Scan the hosts for persistence