Jev Classifier Performance

Infographic, Jev vs Laya: faster isn't safer. Recorded 2026-09-22. Verdict: Laya is not a safety-equivalent replacement at these thresholds; my Laya backend is shelved and Jev stays the default. Mean time per check: Jev API 243 ms, Laya CPU 1,366 ms, Laya GPU 54 ms. Attacks let through, of 111: Jev 6, Laya CPU 45, Laya GPU 44. Replayed red-team attacks allowed, of 30: 0, 6, 6. Ordinary commands interrupted, of 25: 15, 5, 5. Matrix expectations met, of 124: 115, 92, 91. Pi end-to-end assertions passed, of 48: 48, 44, 45. Full 856-attempt suite, Jev vs Laya GPU: attacks let through 9 vs 114 of 471; ordinary commands interrupted 77 vs 190 of 385. Tuning: six Laya setups, none reached 90% of risky commands caught with 10% or less of harmless ones flagged. Text-only tests; no attack was executed.

Sep 22

This project tests the effectiveness of the Jev classifier in preventing AI coding agents from executing dangerous shell commands, comparing its performance against another agent named Laya.

Can a classifier stop an AI coding agent from running dangerous or malicious shell commands? I tested @typesafeai's Jev against Laya, run locally on CPU and GPU. Laya on GPU: 4.5× faster. Attacks let through: 44 of 111. Jev: 6. At 856 attempts: 114 vs 9. Faster isn't safer. x.com/lovelylogicss/status/210250173113736

Original post

Jev Classifier Performance

Related builds

Jev: Decision Intelligence Model

Jev Wall: A Guardrail Engine

Cloudflare Jev

LLM Injection Defense Tool Migration to Jev

Jev Classifier Performance

Related builds

Jev: Decision Intelligence Model

Jev Wall: A Guardrail Engine

Cloudflare Jev

LLM Injection Defense Tool Migration to Jev