Researchers build AI defense that learns from repeated attacks — 79% success rate blocking them

A team created a system where multiple AI agents work together to defend language models against attacks that get smarter across multiple rounds of interaction. Instead of blocking attacks the same way each time, the system remembers what happened before and adapts its defenses, reducing successful attacks by nearly 80% compared to existing defenses.

Language models are now being deployed in real systems where adversaries can attack them repeatedly, refining their approach each round — like a hacker probing a bank vault multiple times. Previous defenses were static, designed to block one attack the same way every time, which doesn't work when attackers learn and shift tactics. This research shows that defenses can themselves learn and adapt across rounds, which means the cost of attacking a deployed language model just went up significantly.