Anthropic created an inspection layer that enables companies to prevent sensitive data from reaching Claude.
TL;DR: Anthropic has introduced inference hooks for Claude Enterprise, which ensures that every prompt and tool call passes through the customer’s DLP server before being processed by Claude. This works with solutions from Netskope, Palo Alto, and Zscaler, and it can be configured for all of Claude’s applications with a single setup.
Anthropic has unveiled inference hooks for Claude Enterprise, which directs all prompts and tool requests through a customer's security server before any processing by Claude. The DLP server reviews the content and provides an approval or denial decision, and Claude will only continue once that decision is made. This same verification occurs with tool calls; when Claude utilizes tools like MCP, skills, or plugins, the tool's responses are verified prior to reaching the model.
This feature addresses a specific enterprise need: security professionals require that any communication channel capable of transmitting sensitive information be inspected by a controlled point. Previously, inline enforcement was confined to client-side hooks in Claude Code. Inference hooks broaden this capability to all aspects of Claude Enterprise, including chat, Claude Code, Cowork, and integrated tools, all managed through a single organisational configuration without needing separate integrations for each product.
The system employs a webhook-driven protocol with a defined schema, allowing it to integrate with existing DLP frameworks from Netskope, Palo Alto Networks, Proofpoint, Zscaler, or custom solutions. Shadow mode (always allows) enables teams to observe before applying restrictions. Role-based exclusions and percentage-based rollouts facilitate gradual implementation. The White House is enforcing control over access to advanced models, and inference hooks serve as Anthropic’s response, allowing companies to manage what their employees communicate with the model.
This feature mitigates concerns raised by figures like Nadella about companies "outsourcing their thinking" to AI providers. With inference hooks, if a company's security server denies a request, Claude never sees it, keeping the data within the enterprise network. Startups are raising substantial funds to develop AI security oversight mechanisms, while Anthropic is integrating that functionality directly into Claude instead of relying on third-party solutions. This feature is currently in beta for Claude Enterprise customers.
Other articles
Anthropic created an inspection layer that enables companies to prevent sensitive data from reaching Claude.
Inference hooks enable security teams to examine each prompt and tool call prior to processing by Claude. They are compatible with Netskope, Palo Alto, and Zscaler. A single configuration applies to all Claude products.
