The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/inspect_eval_uv.py, scripts/inspect_vllm_uv.py, and scripts/lighteval_vllm_uv.py utilize subprocess.run to invoke external command-line utilities (inspect and lighteval). Arguments for these commands are constructed using user-provided input such as model identifiers and task names.
[REMOTE_CODE_EXECUTION]: Both scripts/inspect_vllm_uv.py and scripts/lighteval_vllm_uv.py allow the passing of a --trust-remote-code parameter to the underlying Hugging Face libraries. This enables the execution of arbitrary Python code contained within a model's repository on the Hugging Face Hub during the evaluation process.
[REMOTE_CODE_EXECUTION]: The script scripts/inspect_eval_uv.py performs dynamic path resolution by importing the inspect_evals module and using its installation path as the working directory for shell command execution.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data (model identifiers, task strings, and model outputs) and has exploitable capabilities (shell command execution).
Ingestion points: Model IDs and task specifications provided by the user via CLI arguments; generated outputs from evaluated models (all scripts).
Boundary markers: None. The skill lacks explicit instructions or delimiters to prevent the agent from following commands embedded within the data being processed.
Capability inventory: Execution of system commands through subprocess.run to call inspect and lighteval binaries (all scripts).
Sanitization: While arguments are passed to subprocess.run as a list to prevent shell meta-character injection, the skill does not perform validation or filtering of the model-provided content before it enters the agent's context.

hugging-face-community-evals