The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to utilize the --trust-remote-code flag when serving models. This parameter allows the execution of arbitrary Python code contained within remote model repositories. While a standard feature for custom models, it presents a significant security risk if the model source is not thoroughly vetted.
Evidence: vllm serve MODEL --trust-remote-code in SKILL.md and references/troubleshooting.md.
[COMMAND_EXECUTION]: The troubleshooting guide suggests using administrative privileges to modify system firewall rules, which could lead to unauthorized system changes if the environment is not properly secured.
Evidence: sudo ufw allow 8000 in references/troubleshooting.md.
[DATA_EXFILTRATION]: Deployment instructions suggest binding the inference server to all network interfaces (0.0.0.0) while using a default empty API key. This configuration exposes the service and any data it processes to the network without authentication.
Evidence: vllm serve ... --host 0.0.0.0 and api_key='EMPTY' in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates processing untrusted data which could contain malicious instructions designed to influence the agent or the underlying model.
Ingestion points: Input prompts loaded from prompts.txt and user queries received via the OpenAI-compatible API endpoint.
Boundary markers: No explicit delimiters or instructions are provided to the model or agent to differentiate between instructions and data.
Capability inventory: The skill enables network serving and offline batch generation capabilities across multiple scripts.
Sanitization: There is no evidence of input validation, escaping, or filtering for external content before it is processed by the inference engine.
[DYNAMIC_EXECUTION]: The instructions utilize runtime code execution patterns, including Python one-liners for server testing and dynamic model loading mechanisms.
Evidence: Use of python -c for testing OpenAI SDK integration and dynamic loading of quantized models at runtime.

inference-serving-vllm