The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains instructions for the agent to execute shell commands to install dependencies and manage inference services. Nested bullet points for evidence details: • Commands include pip install vllm, vllm serve meta-llama/Llama-3-8B-Instruct, and pip install locust in SKILL.md.
[PRIVILEGE_ESCALATION]: The documentation advises the user or agent to use elevated privileges to perform network configuration changes. Nested bullet points for evidence details: • The skill suggests executing sudo ufw allow 8000 to open ports in the system firewall within references/troubleshooting.md.
[DYNAMIC_EXECUTION]: The skill frequently encourages the use of a specific vLLM/Transformers flag that enables the loading and execution of arbitrary code from remote sources. Nested bullet points for evidence details: • Instructions in SKILL.md and references/troubleshooting.md recommend using the --trust-remote-code flag for custom models, which permits the execution of Python code bundled with model weights.
[REMOTE_CODE_EXECUTION]: By recommending the --trust-remote-code flag, the skill facilitates a path for remote code execution if the model repository is untrusted or compromised. Nested bullet points for evidence details: • The --trust-remote-code flag triggers the execution of remote scripts from model hubs like Hugging Face.
[INDIRECT_PROMPT_INJECTION]: The skill defines a batch inference workflow where data from a local file is processed by the agent's LLM tools without established safety boundaries. Nested bullet points for evidence details: • Ingestion points: The skill reads input strings from prompts.txt in SKILL.md (Workflow 2, Step 1). • Boundary markers: No delimiters or 'ignore' instructions are provided for the processed file content. • Capability inventory: The skill uses vllm serve and file writing operations (results.jsonl). • Sanitization: There is no logic provided to sanitize or filter the content of prompts.txt before it is passed to the generation engine.

serving-llms-vllm