Hey Product Hunt 👋 I built diplomat-agent after noticing a structural pattern across AI agent codebases. Functions that can charge cards, send emails, and delete data reach production with nothing between the LLM's decision and the real-world consequence. This isn't negligence — framework authors say hard enforcement is the operator's job. In practice, that operator is often nobody. Concrete example from our scan: Khoj's ai_update_memories calls session.delete() + session.add() with no confirmation, no rate limit, no validation. One adversarial prompt wipes a user's memory store. The scanner uses AST (not regex) and is intra-procedural + same-package decorators. Limits are documented — it's a static layer, not a silver bullet. The runtime layer (diplomat-gate) is a separate project. Full methodology and per-repo breakdown in REALITY_CHECK_RESULTS.md. Curious what your agent codebase looks like when you scan it. Happy to dig into specific findings.

diplomat-agent

Find unguarded tool calls in your AI agent code

Find unguarded tool calls in your AI agent code