Anthropic launched inference hooks in beta for Claude Enterprise on 5 August 2026. The mechanism is simple to describe and significant in effect: every prompt, and every tool response flowing back to the model, is sent over a signed WebSocket connection to an organisation’s own security server before Claude processes it. That server returns an allow or deny verdict in real time — a checkpoint that inspects what’s about to reach Claude and can block it before anything sensitive leaves the environment. The coverage spans Claude chat, Claude Code, and Cowork, plus tool calls made through MCP connectors, skills, and plugins, and it plugs into existing DLP platforms — Netskope, Palo Alto Networks, Proofpoint, Zscaler — via a webhook-based protocol. Rollout features include shadow mode for testing without enforcing, role-based exclusions, and percentage-based staged deployment.
So what
This targets the exact objection that stalls AI-assisted development sign-off inside larger, more regulated organisations: “how do we stop an AI coding agent from seeing, or worse, transmitting, something it shouldn’t while it’s got repo and tool access?” Until now the answer has mostly been process — scoped permissions, code review, sandboxing — rather than a real-time technical control sitting in front of the model itself. Inference hooks don’t replace that governance work, but they close a gap in it, and they’re a useful reference point if you’re the one making the case internally for adopting Claude Code or Cowork on client-facing or regulated work. If your organisation is weighing how much autonomy to give AI coding tools and what controls need to sit around them before you can commission that work with confidence, that’s the governance layer we help clients design as part of AI-assisted development engagements — get in touch if you want to talk through what that would look like for your stack.