serving-llms-on-instinct
Pass
Audited by Gen Agent Trust Hub on Sep 10, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local and remote shell commands using subprocess.run to interact with GPU drivers, Docker, and system configurations. This includes using SSH for remote deployments and constructing docker run commands dynamically.
- [EXTERNAL_DOWNLOADS]: The sync_recipes.py script shallow-clones the vllm-project/recipes repository from GitHub and queries Docker Hub for image tags. The estimate_vram.py script fetches model details from the Hugging Face Hub API. These interactions use well-known services to retrieve necessary deployment data.
- [PRIVILEGE_ESCALATION]: Detection and validation scripts use sudo to perform tasks like running amd-smi or adjusting system numa_balancing settings. These actions are part of the intended environment setup for high-performance GPU workloads.
- [INDIRECT_PROMPT_INJECTION]: The skill processes structured data from external sources (GitHub, Hugging Face) that influence the generation of shell commands and vLLM parameters. Ingestion points: Recipe YAML files from GitHub and model config JSON from Hugging Face. Boundary markers: Absent; external values are directly used in command construction. Capability inventory: Execution of shell commands, Docker containers, and SSH operations. Sanitization: Values from external sources are not explicitly sanitized before being used in shell command templates.
Audit Metadata