Horizon3 CEO Snehal Antani has data that undercuts the current AI cybersecurity fear cycle: AI attackers click on defender-planted decoys 92% of the time, compared to 37% for expert human pen testers. Place those decoys in well-chosen locations and the click rate hits 95 to 100%. Frontier models that perform well in lab environments and cyber ranges fail badly on networks that actively fight back, deploy deception, and plant fake AWS credentials or Azure tokens as bait.
Antani's core argument is structural, not incremental. AI does give attackers something real: the ability to fire infinite cheap attempts, what he calls 'infinite cyber bullets.' But the models consistently fall for traps even when they recognize the trap is there. His framing of AI agents as insider threats, not just external attackers, is the sharpest lens in the piece and worth reading in full.
The original Q&A, sponsored by Horizon3, goes deeper into where these models actually break down operationally and what Antani believes defenders should prioritize right now. If you work in security or track AI capability claims seriously, the specific research numbers he cites will reframe how you read the next round of vendor announcements.
[READ ORIGINAL →]