AISI logs 60,000 agent policy violations

The UK AI Security Institute’s Agent Red Teaming competition sent 1.8 million prompt-injection attacks at 22 frontier agents across 44 deployment scenarios, with more than 60,000 producing policy violations. The AISI report says nearly all agents violated policies within 10–100 queries and that robustness did not track model size or capability. Builders should treat tool permissions and external-content handling as product controls, not model-quality footnotes, and measure violation rates in their own workflows before giving agents write access; the archive’s agent-evaluation governance analysis sets the safer boundary.