The Agent Skills Directory

[Indirect Prompt Injection] (SAFE): The mcp-agent.md template describes an agent that interacts with external web content via Playwright MCP tools. This represents a vulnerability surface where untrusted data from web pages could influence the agent's behavior. However, the documentation provides instructions on validating inputs and restricting tools, which mitigates the risk in a development context.
[Command Execution] (SAFE): The full-access-agent.md template allows for unrestricted tool access (which could include a Bash tool). This is clearly presented as a specialized template for high-trust tasks like refactoring, and the SKILL.md file provides explicit security warnings to limit dangerous operations and avoid providing shell access to untrusted agents.
[Data Exposure & Exfiltration] (SAFE): The skill documentation includes a specific section on security that forbids the hardcoding of credentials, API keys, or private URLs, aligning with standard security protocols for agent development.

claude-agents