The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions to use sudo ufw allow 8000 in the troubleshooting guide to modify system-level firewall rules, which requires administrative privileges.
[REMOTE_CODE_EXECUTION]: The documentation repeatedly recommends using the --trust-remote-code flag when serving models (found in SKILL.md and references/troubleshooting.md). This setting enables the execution of arbitrary Python code defined within the model's configuration or modeling files, posing a risk if the model source is not verified.
[COMMAND_EXECUTION]: Instructions in SKILL.md and references/server-deployment.md suggest binding the vLLM server to 0.0.0.0, which exposes the API endpoint to all network interfaces, potentially increasing the attack surface.
[INDIRECT_PROMPT_INJECTION]: The batch inference workflow in SKILL.md demonstrates an attack surface for indirect prompt injection.
Ingestion points: The script in SKILL.md reads input data directly from a local file named prompts.txt.
Boundary markers: There are no boundary markers or instructions to the agent to ignore potentially malicious content within the file.
Capability inventory: The ingested data is passed to llm.generate(), allowing external content to influence the LLM's output and behavior.
Sanitization: The skill does not implement any validation or sanitization for the prompts loaded from the external file.

serving-llms-vllm