Agent Security Harness

Security test AI agent systems against protocol-level attacks including tool poisoning and capability escalation