vllm-prefix-cache-bench
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the official vllm source code from its GitHub repository and downloads a public conversational dataset from Hugging Face for cache efficiency testing.
- [COMMAND_EXECUTION]: Executes Python scripts and vLLM CLI commands to run performance benchmarks and start inference servers.
- [REMOTE_CODE_EXECUTION]: Downloads the official vllm repository and executes internal benchmarking scripts.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external conversational data from the ShareGPT dataset for benchmarking purposes. 1. Ingestion points: Dataset file path provided to the benchmark script in Option 2. 2. Boundary markers: None. 3. Capability inventory: Execution of Python scripts and vLLM CLI tools. 4. Sanitization: None.
Audit Metadata