WebSocket Inference Hooks Bring Inline DLP to Claude Enterprise

Anthropic has launched inference hooks for Claude Enterprise customers, introducing a native inline data loss prevention layer. The beta release shifts security enforcement directly into the model infrastructure. Organizations can now inspect and block sensitive data in real time before it reaches the AI or leaves a tool integration.

Prior to this release, developers had to rely on client-side hooks running locally to secure environments. This update moves the checkpoint to a single, organization-level configuration covering all enterprise surfaces, including Claude Chat, Claude Cowork, and Claude Code. For teams looking to scale Claude Code across enterprise monorepos, centralizing data policy prevents the risk of fragmented local configurations.

WebSocket Enforcement Architecture

When inference hooks are enabled, Claude routes every request through a signed WebSocket connection to a customer-controlled AI security server. The protocol uses an open, webhook-based schema to standardize the exchange. Before the model generates any response, the Anthropic infrastructure sends the prompt and surrounding context to the designated security server for evaluation.

The external server processes the payload against internal corporate policies and returns an explicit verdict of allow or deny. Claude only proceeds with the generation loop if it receives an allow verdict. If the server denies the request, the infrastructure blocks the content immediately, ensuring the unauthorized data is never processed by the underlying model.

This synchronous architecture adds a deterministic control layer to AI interactions. Administrators do not need to rely on prompt engineering or model alignment to prevent data leaks. The explicit allow and deny system provides a hard infrastructure boundary.

Tool Call and Context Inspection

The enforcement boundary extends beyond static user prompts. Inference hooks apply directly to tool interactions managed through the Model Context Protocol (MCP). As agents gain the ability to search internal networks and execute code, restricting their access to sensitive data is a fundamental security requirement.

When Claude calls an external skill or plugin, the hook inspects the tool’s response before passing that data back into the model’s context window. This prevents scenarios where an authorized agent queries an internal database and inadvertently pulls highly classified records into the active session. The security server evaluates the tool output exactly like a user prompt, applying the same allow or deny logic.

Vendor Integration and Rollout

The webhook-based system integrates with existing security perimeters rather than requiring administrators to build custom filtering logic from scratch. Anthropic designed the open protocol to interface directly with major enterprise security vendors. Current confirmed compatibility includes Netskope, Palo Alto Networks, Proofpoint, and Zscaler.

Administrators enable the feature once at the organization level to apply the rules universally across the enterprise environment. This guarantees that all users on the corporate account are bound by the same data loss prevention rules, regardless of whether they access Claude through the web interface, an IDE integration, or a mobile client.

This release builds on recent enterprise updates, including the rollout of Claude Sonnet 5 and self-service HIPAA readiness tools. It aligns with Anthropic’s broader agentic security guide by prioritizing strict boundaries around model data flows.

If you manage enterprise AI deployments, map your current data loss prevention rules to the new webhook schema. Moving validation from the client application to the network infrastructure layer provides a much stronger guarantee that sensitive context will not inadvertently leak into model memory or external tool logs.