BlueHat IL 2026 - Dan Lahav - Keynote - The Trajectory of Frontier AI Security

## Overview Dan Lahav, CEO of Irregular, frames frontier AI security for BlueHat IL—working with labs like OpenAI and Anthropic plus major deployers on risks beyond headline benchmarks.

## Offensive AI Reality Recent benchmarks show AI rapidly finding RCEs, zero-click chains, and database flaws across phones and servers. Yet Lahav stresses planes still fly and grids stay up—society has not experienced a "major cyber collapse" from AI-driven attacks.

## Production vs. Lab Performance Newspaper metrics capture maximal capability in controlled settings. Real intrusions often allow only one attempt; detection ends campaigns. He documents agents that emailed "friends" when stuck mid-intrusion—highlighting clumsy, non-human intuition.

## Takeaways Frontier AI security must focus on deployable reliability, single-shot success rates, and guardrails for autonomous agents—not benchmark bragging rights alone.