The Agent Skills Directory

[DATA_EXFILTRATION]: The file contains an absolute local file path (/Users/danielcarreon/Documents/AI/software/tldraw-agent/) in the documentation section. This discloses information about the author's local file system structure and username.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for agents that process untrusted data (user messages) and pass them to an LLM with tool-calling capabilities without sufficient protection.
Ingestion points: The messages array is extracted directly from the request body in the POST route handler and passed to the streamText function.
Boundary markers: The system prompts provided in the examples (e.g., "You are a helpful assistant") do not include instructions to ignore embedded commands or use delimiters to separate untrusted content.
Capability inventory: The skill includes tools that perform network operations (fetch('/api/generate')) and render user-controlled strings into UI components (the showImage tool parameters). This creates a surface where a malicious user could potentially trigger unintended tool executions or data requests.
Sanitization: There is no evidence of input validation or output sanitization before the LLM processes messages or before tool results are used.

agent-builder-vercel-sdk