I tested my AI safety filter 400 times. It flagged my own innocent prompts 1 in 40

Description and use

The enki‑harness test harness uncovered a 1‑in‑40 false‑positive rate in an AI safety filter, leading to removal of a faulty moderation step.

Problem

AI safety filters often pass tests written by their creators, missing real false positives and negatives.

Outcome

The harness uncovered a 1‑in‑40 false‑positive rate on innocent prompts, leading to removal of a faulty moderation step and saving development time.

Workflow

• 1. Install enki‑harness (MIT, plain Python).
• 2. Run harness locally or as MCP server to register filter with coding agents.
• 3. Execute hb judge to send cases to an open‑weight model (e.g., Llama Guard 3 8B via Ollama).
• 4. Harness generates twins, controls, rules‑off, attacker slice, and falsifier cases.
• 5. Collect flag rates and identify false positives.
• 6. Adjust or remove problematic filter rules based on results.

Evidence

enki‑harness builds the harness for you and adds the parts you'd skip: Twins, Controls, Rules off, Attacker slice, Falsifier

Source

hb judge talks to open‑weight models running locally through Ollama (I used Llama Guard 3 8B). There are no API keys, no per-call cost, and nothing leaves the machine.

Source

Run each one several times and the screen still flagged them at random, about 1 time in 40: a different sentence each time, never the same one twice in a row.

Source