The Agent Skills Directory

[COMMAND_EXECUTION]: The diagnostic scripts vllm_skill_doctor.py and collect_env_summary.py use subprocess.run to execute system commands such as nvidia-smi, rocm-smi, and vllm --help. These executions are limited to specific diagnostic utilities and include timeout protections to prevent hangs.
[REMOTE_CODE_EXECUTION]: The skill facilitates the use of vLLM's trust_remote_code feature, which can execute code from external model repositories. The documentation across the skill and its sub-skills consistently flags this as a high-risk operation, requiring explicit user approval before implementation.
[DATA_EXFILTRATION]: The collect_env_summary.py script gathers environment and package details for troubleshooting. It employs a strict whitelist (SAFE_ENV_KEYS) to ensure that only non-sensitive environment variables are collected, effectively preventing the accidental exposure of API keys or other credentials.
[INDIRECT_PROMPT_INJECTION]: The skill provides an interface for processing user-supplied data, including model configurations and multimodal payloads. It includes dedicated validation tools (validate_structured_request.py and validate_multimodal_payload.py) to verify the structure and safety of these inputs, mitigating risks associated with malformed or malicious data processing.
[PRIVILEGE_ESCALATION]: Filesystem and network access for multimodal models are managed through allowed_local_media_path and allowed_media_domains parameters. The skill documents these as essential safety boundaries to prevent unauthorized directory traversal or SSRF attacks.

vllm