Wavestone finds 17% secure agent tools

Only 17% of the 30 organizations in Wavestone’s 2026 AI Cyber Benchmark had incorporated secure access to AI functions and tools into development standards, even as 33% had put agentic AI into governance frameworks. Wavestone’s benchmark says the panel is improving at governance while monitoring, investigation, and response remain comparatively immature. For builders, agent permissions need an enforceable control plane—not another approval document; our Claude Code control-plane analysis makes the same case.