I built a 0.07ms Python firewall to stop AI hallucinations

Name: I built a 0.07ms Python firewall to stop AI hallucinations
Availability: InStock
Author: richardewing1

by richardewing1·Mar 23, 2026·4 points·2 comments

Visit Project View on HN

AI Analysis

●●SolidSolve My ProblemSlick

Deterministic policy gates beat LLM guardrails when your agent tries to DROP TABLE.

Strengths

•0.07ms median latency with zero false negatives across 5000-payload red-team test.
•Eight specific policy gates block rm -rf, subprocess, untrusted API calls, and destructive SQL.

Weaknesses

•AI governance category is getting crowded—Guardrails AI, LaGuardia, and others already exist.
•Pure Python may not suit high-throughput production where Go or Rust firewalls dominate.

Post Description

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