The most urgent AI security threat is the compromise of autonomous agents, creating new attack surfaces that require zero-trust identity and behavioral monitoring.