claude-agent-sdk
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The
scripts/check_drift.pyutility fetches documentation from official sources and GitHub repositories to verify the consistency of the skill's reference material. The script uses therequestslibrary to retrieve data for hashing purposes but does not execute the fetched content. - [COMMAND_EXECUTION]: Files such as
examples.mdandpatterns.mdprovide Python templates for implementing tools that interact with the system shell using thesubprocessmodule. These are provided as educational examples for building agent capabilities. - [DYNAMIC_EXECUTION]: The documentation includes patterns for tools that execute dynamically generated Python or Bash code. These examples are accompanied by guidance on using security hooks to validate and restrict execution parameters.
- [INDIRECT_PROMPT_INJECTION]: The skill documents the attack surface for indirect prompt injection and provides mitigation strategies. 1. Ingestion points: Untrusted data enters the agent context through tool outputs like
web_searchandread_fileas defined inexamples.md. 2. Boundary markers: The skill suggests using subagents for context isolation inpatterns.md. 3. Capability inventory: The skill defines tools withsubprocessexecution and file-system writes inexamples.md. 4. Sanitization: Examples forSecurityValidationHookandInputSanitizationHookare present inpatterns.mdto filter dangerous content from tool inputs.
Audit Metadata