The Agent Skills Directory

Dynamic Execution (MEDIUM): The mcp__gemini__ask-gemini tool documentation explicitly identifies a sandbox: true option for code execution. This allows the agent to execute code generated by the LLM at runtime. Although intended for architectural validation and trade-off analysis, executing dynamically generated code is an inherent security risk.
Indirect Prompt Injection (LOW): The skill is configured to ingest data from the local environment using Read, Grep, and Glob tools to provide context for the AI.
Ingestion points: Local filesystem content read via standard MCP tools.
Boundary markers: The provided prompt templates do not utilize delimiters or specific instructions to ignore embedded commands in the context data.
Capability inventory: The agent has the ability to execute code (via the sandbox tool parameter) and access the filesystem.
Sanitization: There is no evidence of sanitization or filtering of the file content before it is interpolated into the model's prompt.

consulting-design