claude-agent-sdk
Warn
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill provides templates (
templates/custom-tool-python.pyandworkflows/add-custom-tools.md) that demonstrate how to create custom tools using the Pythoneval()function to process mathematical expressions. Although the code attempts to restrict execution by clearing__builtins__, this is an insecure practice in Python that can often be bypassed to achieve arbitrary code execution. Since these tools are designed to be called by an AI agent that may be processing untrusted data, this presents a significant execution risk. - [INDIRECT_PROMPT_INJECTION]: The skill describes and provides code for research-oriented agents (
templates/research-agent-python.pyandreferences/best-practices.md) that ingest content from external websites via theWebFetchtool. These agents are also granted powerful system capabilities, including theBash,Write, andEdittools. The provided examples lack explicit sanitization of the fetched content or the use of strict boundary delimiters, creating a potential vulnerability where instructions embedded in remote web pages could influence the agent to perform unauthorized file or system operations. - [REMOTE_CODE_EXECUTION]: The documentation references the installation and use of official packages and CLI tools from
@anthropic-ai, such as the Claude Agent SDK and Claude Code CLI. These are established tools from a well-known service and are documented here for legitimate integration purposes.
Audit Metadata