claude-agent-sdk

Pass

Audited by Gen Agent Trust Hub on Sep 18, 2026

Risk Level: SAFEDYNAMIC_EXECUTIONDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The documentation and examples include patterns for creating custom tools that use the eval() function to process mathematical expressions. While provided with security warnings to use safe evaluation libraries in production, the inclusion of eval() represents a dynamic execution surface.
  • Evidence: Found in references/custom-tools.md and references/api-reference.md within calculator tool examples.
  • [DYNAMIC_CONTEXT_INJECTION]: The skill documents the use of the !command syntax for slash commands, which allows for the execution of shell commands (such as git status or git diff) and the inclusion of their output into the prompt context at runtime.
  • Evidence: Examples provided in references/slash-commands.md demonstrate how to use this feature for git workflows.
  • [INDIRECT_PROMPT_INJECTION]: The skill provides instructions for building agents that ingest untrusted data (user prompts and codebase files) and have access to powerful tools like Bash, Write, and Edit, establishing an indirect prompt injection surface.
  • Ingestion points: query() and ClaudeSDKClient.query() methods in all example scripts (e.g., examples/quick_start.py, examples/streaming_mode.py).
  • Boundary markers: The skill encourages the use of the claude_code preset but does not explicitly instruct on specific sanitization for interpolated data.
  • Capability inventory: Broad tool access including Bash, Write, Edit, and Task (delegation) tools across various examples.
  • Sanitization: No explicit sanitization or filtering logic is provided for external inputs before they are processed by the SDK client.
  • [COMMAND_EXECUTION]: Multiple examples demonstrate the use of the Bash tool to execute shell commands for tasks like project setup, testing, and file operations.
  • Evidence: See examples/basic-orchestrator.py and examples/streaming_mode.py.
  • [EXTERNAL_DOWNLOADS]: The skill examples and project templates use the uv tool to manage dependencies, including the official claude-agent-sdk package.
  • Evidence: Dependency blocks in examples/basic-orchestrator.py and assets/sdk-template.py.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 18, 2026, 12:53 AM