The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides template files (workflows/add-custom-tools.md and templates/custom-tool-python.py) that demonstrate using the Python eval() function to process user-provided strings for a calculator tool. Executing untrusted code via eval is a high-risk pattern that can be exploited for arbitrary code execution if not strictly controlled.
[COMMAND_EXECUTION]: The SDK features a built-in Bash tool that allows agents to execute arbitrary terminal commands. Multiple documentation files and templates (references/built-in-tools.md, templates/agent-with-tracking-python.py) show how to enable and use this capability, which provides a direct path to host system manipulation. The SDK documents various permission modes like 'bypassPermissions' that skip safety checks.
[PROMPT_INJECTION]: The skill documentation describes building agents that ingest untrusted data from the web (WebSearch, WebFetch) or local files (Read). This creates a significant surface for indirect prompt injection.
Ingestion points: Files like references/built-in-tools.md describe tools that read external content into the agent context.
Boundary markers: Explicit delimiters or warnings to ignore embedded instructions are not enforced by default in the provided templates.
Capability inventory: The skill includes powerful tools such as Bash, Write, Edit, and Task (references/built-in-tools.md).
Sanitization: The SDK provides hooks and permission handlers for sanitization (references/permissions.md), but they require manual implementation by the developer.

claude-agent-sdk