creating-mcp-agent-evals
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides instructions for running various development and maintenance commands through the terminal. Evidence includes commands in
reference.mdfor executing evaluation suites withpnpm, managing remote datasets vianpx langfuse-cli, and running local probe scripts withtsx.\n- [INDIRECT_PROMPT_INJECTION]: The evaluation framework described in the skill ingests user-provided text prompts to test agent behavior, creating a surface for indirect prompt injection. \n - Ingestion points: The
input.queryfield in the Langfuse dataset item shape described inreference.md. \n - Boundary markers: No explicit boundary markers or instruction-override directives are specified for the query input. \n
- Capability inventory: The harness executes MCP tool calls and can run local shell commands. \n
- Sanitization: No specific sanitization or filtering of the user queries is mentioned, as the purpose is to evaluate model responses to arbitrary natural language. \n- [SAFE]: The skill follows standard security practices by instructing users to store credentials in
.envfiles rather than hardcoding them. All external domains and repositories mentioned are either vendor-owned (e.g., langfuse.apify.dev) or recognized technology services (e.g., rfc-editor.org, example.com).
Audit Metadata