The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The specialists use web_search and web_fetch to research topics, which introduces a surface for indirect instructions to influence agent behavior.
Ingestion points: Research tools in subagents/podcast-producer/instructions.md and others.
Boundary markers: None identified in instructions to delimit untrusted web content.
Capability inventory: File writes to Notion, email sending via Resend, social posting via Typefully, and file uploads to Vercel Blob.
Sanitization: No explicit sanitization or validation of retrieved web content is documented.
[DATA_EXFILTRATION]: The uploadAssetTool in lib/vercel-blob/upload-asset-tool.ts enables the agent to upload files to public storage.
The tool converts base64 input to public URLs via Vercel Blob.
This functionality, while useful for sharing assets, could be leveraged to expose sensitive data if the agent is compromised.
[EXTERNAL_DOWNLOADS]: The skill fetches code and configurations from external sources.
Downloads the template repository from Vercel Labs' official GitHub account.
Fetches and runs Model Context Protocol servers for Notion and Typefully via npx.
[COMMAND_EXECUTION]: The documentation and integration code involve executing shell commands.
Installation instructions use npm install and git clone.
The Notion connection in connections/notion.ts executes npx -y @modelcontextprotocol/server-notion to start the integration server.

vercel-marketing-team-eve-template