claude-agent-sdk
Pass
Audited by Gen Agent Trust Hub on Sep 26, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The file 'templates/custom-mcp-server.ts' includes a demonstration tool ('calculate') that uses the JavaScript 'eval()' function to process input expressions.
- Evidence: The tool handler implementation contains 'const result = eval(args.expression);' which executes a string provided by the agent.
- Context: This implementation is part of an educational template and is accompanied by a comment recommending the use of a secure math parser in production.
- [INDIRECT_PROMPT_INJECTION]: The skill provides instructions for building agents that ingest and process external data, creating a potential surface for indirect prompt injection.
- Ingestion points: The documentation in 'SKILL.md' and the MCP server guides describe fetching content from external APIs.
- Boundary markers: The skill promotes the use of JSON schema validation through the 'outputFormat' option to restrict agent responses.
- Capability inventory: Templates provide examples using powerful capabilities like 'Bash', 'Write', and 'Edit' across various scripts including 'templates/multi-agent-workflow.ts'.
- Sanitization: 'SKILL.md' (Issue #14) includes specific remediation guidance for sanitizing Unicode line separators in external data to prevent parsing errors.
Audit Metadata