mcp-builder
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill implements utilities to execute local commands for the purpose of launching and interacting with MCP servers during development and testing.
- Evidence:
scripts/connections.pyandscripts/evaluation.pyutilize themcp.client.stdiotransport to manage server subprocesses. - [INDIRECT_PROMPT_INJECTION]: The automated evaluation harness ingest results from external tools, creating a vulnerability surface where a malicious tool output could attempt to override the evaluating agent's instructions.
- Ingestion points:
scripts/evaluation.pyreads tool results in theagent_loopfunction and injects them into the message history. - Boundary markers: The
EVALUATION_PROMPTrequires the agent to wrap its output in specific XML tags (<summary>,<feedback>,<response>), providing a structural defense against output hijacking. - Capability inventory: The skill can execute shell commands via the
mcpSDK and perform network operations via the Anthropic API. - Sanitization: Tool outputs are stringified and included in the prompt context without content-based filtering.
Audit Metadata