How the attack works
An attacker publishes a normal-looking 'frontend' model repository whose config.json contains an auto_map entry pointing to a separate 'backend' repository they control. When vLLM loads the model, it resolves that auto_map reference with get_class_from_dynamic_module() and fetches the backend repository's Python file, bypassing the trust_remote_code=False setting that should have blocked custom code. That fetched module then executes inside the vLLM process. The detection watches one process tree at a time: it flags the required step of a dynamic Python module being pulled from a different repository than the original config, and corroborates it either by seeing the config-fetch-then-redirect pattern or by the same session immediately spawning a shell, network tool, encoder, or interpreter payload.
Why it matters
An attacker can get arbitrary code execution inside the model-serving process without the operator ever setting trust_remote_code=True on the malicious repository, potentially leading to data theft, credential access, or further lateral movement from the serving host.
What you can do
- Pin and vet every repository referenced directly or via auto_map before serving a model; treat auto_map redirects to a different repo as a red flag requiring manual review.
- Run vLLM model-loading processes with least privilege and no access to credential files or internal network paths they don't need.
- Monitor for a model-loading process that fetches a .py file from one repository and then spawns a shell, network utility, or encoded payload shortly after.
- When trust_remote_code is required, restrict it to a known, pinned repository and reject configs whose auto_map points outside that repository.
Known benign look-alikes
- Legitimate trust_remote_code=True workflows where a model's config.json auto_maps to a shared base repository (some multimodal and Nemotron-family checkpoints do this by design). This produces the cross-repo stage without any malicious intent - triage by confirming the backend repository owner.
- Model conversion, quantisation or evaluation pipelines that pull remote modelling code and then legitimately shell out (chmod, bash helper scripts, pip install of a missing kernel package) inside the same Python process tree within the 10 minute window.
- Interactive notebook or dev sessions where an engineer downloads a modelling_*.py from a hub UI and separately runs curl/ssh from the same Python parent.
- CI image builds that fetch model code and post-process it with shell tooling under one python entrypoint.