The Agent Skills Directory

REMOTE_CODE_EXECUTION (HIGH): The skill instructs the user to execute a remote script using curl -fsSL https://ollama.ai/install.sh | sh.
Evidence: Found in the 'Ollama (Local Deployment)' section of SKILL.md.
Reasoning: Executing piped web content is a significant security risk as the source (ollama.ai) is not within the defined trusted organization scope and the content could be modified to execute malicious code on the host system.
COMMAND_EXECUTION (MEDIUM): The skill provides multiple instructions for running arbitrary commands and services with high privileges.
Evidence: docker run --gpus all and python -m vllm.entrypoints.openai.api_server.
Context: These are necessary for the skill's primary purpose of model deployment, but they involve running external images and starting network-accessible servers.
INDIRECT_PROMPT_INJECTION (LOW): The FastAPI implementation creates a surface for indirect prompt injection.
Ingestion points: The GenerateRequest.prompt field in the FastAPI server (file: SKILL.md).
Boundary markers: None present in the code; the prompt is passed directly to the tokenizer.
Capability inventory: The model generates text based on the input, which can be used to influence downstream agent behavior if the output is processed further.
Sanitization: No input validation or sanitization is performed on the prompt string before inference.

model-deployment