As AI agents gain autonomy, the attack surface expands. Red-teaming methodologies, jailbreak prevention, agent-specific security controls, and kill switches.