The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's orchestration scripts, specifically scripts/auto_optimize.py, use subprocess.run(shell=True) to execute benchmark commands constructed from concatenated strings. Other scripts like scripts/run_bench.sh and scripts/run_batch.sh utilize bash -c for similar purposes. This practice of executing shell commands built from external or untrusted parameters without proper escaping is a known security risk.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting metadata from a remote inference service and using it in shell commands. In scripts/probe_service.sh, the skill fetches model identifiers and weight paths from the service's /v1/models endpoint. These values are used directly as command-line arguments in subsequent benchmarking phases without sanitization.
Ingestion points: Data enters the skill context through the service discovery process in scripts/probe_service.sh which queries the target's /v1/models API.
Boundary markers: There are no boundary markers or instructions to isolate or ignore malicious content within the service-provided strings.
Capability inventory: The skill possesses extensive command execution capabilities through its various shell and python scripts.
Sanitization: No sanitization, escaping, or validation of the model IDs or paths is performed before they are interpolated into shell commands.
[EXTERNAL_DOWNLOADS]: The skill is designed to download datasets from HuggingFace Hub, which is a common and established service for hosting machine learning datasets.

vllm-bench-serve