The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/run_bench.sh, scripts/run_batch.sh, and scripts/auto_optimize.py build shell commands for vllm bench serve by interpolating variables provided by the user, such as base_url, model, and dataset_path. Without strict validation, this pattern is susceptible to command injection.
[DYNAMIC_EXECUTION]: In scripts/auto_optimize.py, the skill executes shell commands using subprocess.run(cmd, shell=True). This is a known risky pattern when the command string contains dynamic content derived from external or user-supplied sources.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect injection as it ingests data from user-specified endpoints and file paths which are then leveraged in shell execution. \n
Ingestion points: User-supplied parameters such as base-url, model, and dataset-path used throughout the benchmarking workflow. \n
Boundary markers: The skill does not implement boundary markers or instructions to the agent to ignore embedded commands in the processed data. \n
Capability inventory: The skill possesses the ability to execute shell commands and interact with network services. \n
Sanitization: There is no evidence of sanitization or escaping of the input variables before they are interpolated into shell command strings.

vllm-bench-serve