Anthropic ships Inference Hooks so security teams can gate Claude prompts and tool calls

When Inference Hooks are on, Claude routes each inference request through a signed connection to the organisation's security server. Before generation starts, Claude sends the prompt and surrounding context; the server returns allow or deny, and Claude only proceeds on an allow verdict[2][3]. The same pattern applies to tool calls: prompt hooks check what people send in, and pre-tool-call hooks check tool calls in supported products before they run[1].

Anthropic says the open webhook-style protocol lets teams reuse policies already written for Netskope, Palo Alto Networks, Proofpoint, Zscaler or an in-house AI security server[2]. Rollout helpers include shadow mode (always allow), role-based exclusions and percentage-based rollouts, plus configurable timeouts and failure-policy tolerance[2].

Turning Inference Hooks on at the organisation level applies them across Claude Enterprise surfaces, including chat, Claude Code, Claude Cowork, and tool calls made through MCP connectors, skills and plugins[1][2]. Bandwidth's vice president of information security, Andrew Grimmett, said the hooks add a checkpoint to inspect what is flowing to Claude in real time before the model sees it, letting the company move faster on AI without giving up control[2].

Inference Hooks arrive alongside a wider Claude Enterprise security update: a security center that compares live settings with Anthropic's recommended configuration, centralized management for parent organisations at admin.claude.com, Admin API member sign-out, self-serve IP allowlists, Claude Desktop controls that can block personal accounts on managed machines, and an optional EDR sensor inside the local Cowork virtual machine[1]. Anthropic notes that the security center and several access features are still in beta, while Inference Hooks are generally available[1].

Enterprise AI adoption often stalls on the same question: can the security team apply existing DLP and tool policies to every chat and agent surface without a bespoke integration per product. Inference Hooks are Anthropic's answer for Claude Enterprise: keep the policy server you already trust, and make Claude wait for its verdict[1][2].