Igor Kozlov

Despite excelling on curated security Q&A benchmarks, today's LLMs fail spectacularly at open-ended threat hunting, correctly identifying less than 4% of malicious events in a new rigorous benchmark.