improving-mcp-tools
Pass
Audited by Gen Agent Trust Hub on Aug 4, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill implements a robust 'measure-fix-validate' loop with explicit guardrails. It restricts file modifications to a specific allowlist (
products/*/mcp/tools.yaml,products/*/skills/**, etc.) and mandates the use of an evaluation harness (services/mcp/evals/) to justify any changes. - [COMMAND_EXECUTION]: The skill uses local developer commands (
pnpm dev:hono,pnpm exec tsx) and PostHog-specific analytics tools (query-mcp-tool-stats,query-mcp-tool-failures). These are standard development operations within the PostHog environment and do not involve arbitrary shell injection or untrusted remote execution. - [DATA_EXFILTRATION]: While the skill interacts with local and development databases (
POSTHOG_API_BASE_URL=http://localhost:8000), it explicitly mandates a 'read-only' posture against production data and restricts testing to seeded local/devbox stacks. No non-whitelisted external domains or exfiltration patterns were detected. - [PROMPT_INJECTION]: The instructions contain phrases like 'guardrails below are hard rules' and 'violating any of them ends the campaign pass.' These are structural instructions for the agent's workflow and do not attempt to bypass safety filters or override system-level constraints.
Audit Metadata