Anthropic adds a pre-inference enterprise hook

Anthropic put 1 beta inference hook across three Claude Enterprise surfaces—claude.ai, Cowork, and Claude Code—so each governed prompt can wait for an external security server’s allow-or-deny verdict before reaching the model. The inference-hook documentation sets a five-second default verdict timeout and supports shadow mode plus percentage rollouts; security teams can now enforce policy upstream of generation, complementing the runtime controls emerging inside agent harnesses rather than relying only on after-the-fact audit logs.