The Agent Skills Directory

[COMMAND_EXECUTION]: The troubleshooting guide in references/troubleshooting.md suggests using sudo ufw allow 8000 to open firewall ports. This command uses administrative privileges to modify the system's security configuration, potentially increasing the network attack surface.\n- [REMOTE_CODE_EXECUTION]: Instructions in SKILL.md and references/troubleshooting.md recommend using the --trust-remote-code flag to support custom models. This flag allows the execution of arbitrary Python code included in the model repository, which can lead to remote code execution if a model is loaded from an untrusted source.\n- [DATA_EXFILTRATION]: Deployment patterns in SKILL.md and references/server-deployment.md advise binding the server to 0.0.0.0 and exposing a Prometheus metrics endpoint on port 9090. This configuration exposes the inference API and system performance data to the network, which may lead to unauthorized data access if not properly restricted.\n- [EXTERNAL_DOWNLOADS]: The skill involves the installation of several third-party libraries including vllm, locust, autoawq, auto-gptq, and flash-attn. While these are recognized tools, they represent an external dependency chain.\n- [PROMPT_INJECTION]: The skill facilitates the serving of LLMs that process arbitrary user input, creating a surface for indirect prompt injection.\n
Ingestion points: User prompts are ingested via the OpenAI-compatible API described in SKILL.md.\n
Boundary markers: Absent; there are no instructions for using delimiters or warnings to ignore instructions embedded in data.\n
Capability inventory: The vLLM server has network access and can perform code execution if the trust flag is enabled.\n
Sanitization: Absent; no methods for input filtering or validation are documented.

serving-llms-vllm