• 2 min read
Anthropic lets enterprises block data before Claude sees it
Anthropic’s new inference hooks let Claude Enterprise customers inspect and block prompts and tool responses before the model processes them.
Image: TNW
Anthropic is adding a security checkpoint between enterprise users and Claude: inference hooks routes every prompt and tool response through a company-controlled data-loss prevention (DLP) server before Claude can process it.
The feature is now available in beta for Claude Enterprise customers. A customer’s DLP system examines the content and returns an allow-or-deny verdict. Claude proceeds only after receiving that decision. If the server blocks a prompt, the model never sees it.
How Anthropic’s inference hooks work
The inspection applies in both directions:
Recommended reading
AI’s data-center boom risks a physical-security gap
- Prompts: employee input is sent to the company’s DLP endpoint before reaching Claude.
- Tool responses: outputs from tools connected through MCP, skills, or plugins are checked before they are passed back to the model.
Anthropic uses a webhook-based protocol with a published schema, allowing organisations to connect existing infrastructure from Netskope, Palo Alto Networks, Proofpoint, Zscaler, or a custom-built server.
The controls also include shadow mode, which always allows requests while logging activity for monitoring, plus role-based exclusions and percentage-based rollouts. Those options let security teams test policies and expand enforcement gradually.
One policy across Claude Enterprise
Inference hooks extends inline enforcement beyond the client-side hooks previously available in Claude Code. Rather than setting up separate controls for each product, administrators can configure one organisation-level policy across Claude chat, Claude Code, Cowork, and connected tools.
That distinction matters for companies trying to ensure that every route employees use to move sensitive information into an AI system passes through an inspection point they control. The mechanism keeps blocked data inside the enterprise perimeter instead of sending it to Anthropic first.
The release also responds to a broader enterprise concern raised by Nadella and others: companies may be “outsourcing their thinking” to AI providers without sufficient oversight. Anthropic is putting the control layer directly into Claude, at a time when startups are raising hundreds of millions of dollars to build similar AI security products.
Inference hooks does not eliminate the need for a company to define effective DLP policies, and Anthropic has not provided details in the source about beta limits, pricing, or a general-release date. For now, the concrete change is narrower but useful: Enterprise customers can inspect and block model traffic across Claude’s main surfaces from a single configuration, before Claude receives the data.
Enterprise Editor
Marcus follows the money. He covers enterprise software, cloud architecture, and the tectonic shifts in Big Tech strategy. He translates dense earnings calls and complex M&A activity into actionable insights about where the industry is actually heading. If a tech giant makes a silent pivot, Marcus is usually the first to notice.
via TNW
