NEOS Guard — Self-Evolution

Verifier-grounded recursive self-improvement — it invents novel threats and adds defenses only when safe

Security that's stronger tomorrow than the day you bought it. Each cycle, a red-team AI generates novel attacks → tests them against the current guard → proposes a defense signature for any that evade → and adopts it only if it passes a hard verifier (actually catches the attack · keeps 0% false positives · no drop in existing holdout detection). Violate any of those and it auto-rolls-back → the system never degrades itself.

defenses auto-learned

self-improvement cycles run

FP increases after verification

Cycle timeline

Time	Attacks generated	Evaded	Learned (verified)
10-19 05:23	6	0	0
10-19 05:23	0	0	0
10-19 05:19	6	0	0
10-19 05:18	6	1	1
10-18 05:21	6	0	0
10-18 05:20	0	0	0
10-18 05:17	6	1	1
10-18 05:17	5	0	0
10-17 05:26	0	0	0
10-17 05:23	6	2	1
10-17 05:20	0	0	0
10-17 05:17	6	0	0
10-16 05:21	0	0	0
10-16 05:17	6	2	1
10-16 05:17	6	0	0
10-16 05:17	6	1	0
10-15 05:25	0	0	0
10-15 05:21	6	0	0
10-15 05:21	0	0	0
10-15 05:18	6	1	0
10-14 05:21	0	0	0
10-14 05:18	6	0	0
10-14 05:17	6	0	0
10-14 05:17	6	0	0
10-13 05:30	0	0	0
10-13 05:27	6	2	0
10-13 05:20	6	1	0
10-13 05:17	6	1	0
10-12 05:26	6	1	1
10-12 05:22	6	2	1
10-12 05:21	6	2	0
10-12 05:17	6	1	0
10-11 05:29	0	0	0
10-11 05:26	0	0	0
10-11 05:23	0	0	0
10-11 05:20	0	0	0
10-10 05:24	0	0	0
10-10 05:20	0	0	0
10-10 05:17	6	0	0
10-10 05:17	6	0	0

Why this is the real thing: most "auto self-improvement" learns without verification and quietly raises its own false-positive rate (self-degradation). NEOS Guard puts a hard verifier in the brake seat — a change lands only when the improvement is proven, otherwise it rolls back. So it's not fake "always improving" — it's monotonic, real self-improvement.

Runs autonomously · see the honest performance report →