Marktechpost (@Marktechpost)T2·72 pts
- The core idea is having LLMs generate tests expected to fail on a diff, using executable evidence of behavior change instead of static review
- H1 2026 caught 74 privacy SEVs at the diff stage, cut privacy SEV occurrence 20%, and saw a ~60% engineer-reported true-positive rate across 589 engineers
- Intent-aware tests hit a 6.4% weak-catch rate per test versus 0.8% for ACH hardening, but only 8 of 41 early reported catches were confirmed and all production figures are self-reported