inference-serving-tensorrt-llm
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides instructions and code for building LLM inference services that ingest untrusted user input.
- Ingestion points: The
llm.generate()method inSKILL.mdand the/v1/chat/completionsAPI endpoint documented inreferences/serving.mdaccept arbitrary text input. - Boundary markers: The provided examples do not demonstrate the use of delimiters or specific instructions to mitigate the risk of embedded instructions within user-provided prompts.
- Capability inventory: The skill enables high-performance inference and network-based serving of model outputs.
- Sanitization: No sanitization or validation logic for input prompts is included in the examples.
- [EXTERNAL_DOWNLOADS]: The skill's serving utility (
trtllm-serve) is designed to automatically download model weights from Hugging Face. - Evidence: The command
trtllm-serve meta-llama/Meta-Llama-3-8BinSKILL.mdand thereferences/serving.mdfile describe the automatic fetching of model artifacts from Hugging Face repositories.
Audit Metadata