FlawDetector blog

Field notes from the detection engine

How the Adversarial Patch Loop verifies a fix, what 1,000 repeated trials told us about variance, and why false positives are an operating problem. Written by the engineers who ship the engine.

Latest

EngineeringJul 20269 min read

How the Adversarial Patch Loop decides when a vulnerability is actually sealed

How FlawDetector's AI red team vs blue team loop uses three gates — original exploit, 12 mutations, regression parity — to decide a vulnerability is sealed.

Read article

BenchmarkJun 20268 min read

FlawDetector-LLM V1.0: 1,000 repeated trials, and what the variance told us

94.7% detection, 3.2% false positives, 99.1% verdict consistency over 1.2M verdicts — where the variance lived, and the three changes that closed it.

Read article

SecurityMay 20267 min read

False positives are a culture problem before they are a model problem

A 3.2% false positive rate is 138 findings a week on 2M lines. Separate the four things people call a false positive, then use expiring suppressions.

Read article

See the loop run on your own repository

Connect a repo and get a full scan, a merge-ready patch, and a verdict log for every critical finding. No card required.

Start a free scan Talk to a security engineer