New BrowseSafe system achieves 90.4% accuracy detecting prompt injection attacks in AI browser agents, outperforming existing models with sub-second response times.