The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes vllm bench serve commands within a Docker container to measure performance metrics such as throughput and latency. These operations are restricted to the local environment and the target container.- [DYNAMIC_EXECUTION]: Python scripts scripts/run_benchmark.py and scripts/run_all_benchmarks.py use the subprocess.run() module to interface with system binaries. The command construction uses lists, which is a safe practice against shell injection.- [INDIRECT_PROMPT_INJECTION]: The skill represents a vulnerability surface for indirect prompt injection as it parses potentially untrusted output from models and benchmarking tools. However, it implements standard JSON parsing and regex-based metric extraction which limits the risk of adversarial data influence.- [REMOTE_CODE_EXECUTION]: A pattern involving curl piped to python3 was identified in SKILL.md. Technical review confirms this is a benign operation used to parse JSON data from a local service (localhost:8000) using a static Python one-liner to extract model IDs.

perf-test-flagos