Show HN: I built a 0.07ms Python firewall to stop AI hallucinations

[richardewing1]

Hey HN,

I was building other AI products and kept hitting the same wall: agents hallucinate and constantly forget context. I ended up building a ledger just to track what they were actually doing.

Once that ledger was logging everything, I realized I could use it to enforce rules. Exogram is the result. It is a deterministic firewall that intercepts agent tool calls (like MCP) using pure Python logic gates. It catches bad schemas and destructive actions in 0.07ms. There is zero AI in the security loop.

I load-tested it yesterday and hit 137 RPS on a single container with zero dropped payloads.

Sandbox: https://exogram.ai/proving-ground Repo: https://github.com/Richard-Ewing/exogram-protocol-rfc

I just built the Proving Ground UI today to visualize the 0.07ms block in real time. Let me know if the split-screen demo makes sense, or if you manage to bypass the gates.

Richard

[dhedegreen]

This is the missing layer nobody talks about. We kept connecting AI to real systems - browse this, call that, send this and each step felt like a feature. Nobody stopped to say: wait, the model is no longer just generating text, it's touching things. Cool to see someone building the gate with plain code instead of more AI on top of AI.

[richardewing1]

Yeah exactly, messing up a chat response is whatever, but the second they have actual write access you just can't trust a system prompt anymore.