The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes documentation and samples for agents with high-capability tools that execute model-generated content. Examples include:
Ingestion points: The messages and prompt parameters in files such as samples/natural-language-postgres.md and samples/computer-use-anthropic-tools.md ingest user-controlled text.
Boundary markers: Many snippets omit explicit boundary markers (e.g., delimiters for untrusted input) as they focus on API usage demonstrations.
Capability inventory: The samples demonstrate tools that execute raw SQL queries (samples/natural-language-postgres.md) or shell commands (samples/computer-use-anthropic-tools.md, samples/agent-skills-load-skill-tool.md).
Sanitization: The skill provides explicit security warnings within the documentation, recommending read-only database connections, dedicated VM isolation, and human-in-the-loop approval flows for sensitive operations.
[DYNAMIC_EXECUTION]: Multiple code samples demonstrate the implementation of tools that perform dynamic execution to showcase the library's capabilities. Specifically:
samples/computer-use-anthropic-tools.md shows a bash tool using child_process.execSync to run model-supplied commands.
samples/agent-skills-load-skill-tool.md implements a tool that executes commands within a provided sandbox abstraction.

vercel-ai-sdk