By the time a security team reviews a risky prompt in a log, the model has already answered it. Anthropic's new Inference Hooks close that gap for Claude Enterprise: every governed prompt and tool result on claude.ai, Cowork, and Claude Code can now be checked by an organization's own AI security server at runtime, before inference runs, not after. Island’s Enterprise Agentic Control Plane can now act as that server. 

Inline AI security, at runtime

Island connects to Inference Hooks as that security server, routing every governed prompt and tool result straight into Island's policy engine, the Enterprise Agentic Control Plane, the same layer already enforcing DLP and audit on Claude traffic elsewhere. Nothing new to install - no browser, no extension, no MDM profile. The connection is server-to-server, so the same two-layer enforcement the Control Plane already runs on Claude traffic, deterministic DLP and an LLM judge, evaluates the request before Claude responds. A denial happens inline, with the reason surfaced to the user and the event logged in the Control Plane's audit trail alongside everything it already sees on the browser, network, and endpoint.

The hook itself is vendor-neutral. Anthropic built it to work with whatever security server an organization already points its other tools at, so plugging in a verdict is no longer the hard part. What differs is what stands behind that verdict. A network-only view sees the transcript in isolation. Namely, one prompt, one decision, no history. Island's verdict carries everything the Control Plane already knows about that user in that moment, the browser session, the endpoint, the MCP calls before and after, all under one identity. The prompt never gets judged alone.

Inference Hooks is Anthropic's first inline checkpoint for Claude Enterprise. For Island, it's a new vantage point in AI inference security, next to the browser, network, endpoint, and the Claude Compliance API integration, judging the prompt in the same identity and policy context the Control Plane already holds for that user, before the model ever sees it.

Get started

Island's Inference Hooks integration connects directly to the Enterprise Agentic Control Plane, with no additional footprint to deploy. Available to Island customers today.