Skip to content
NewsProduct launch

Anthropic launches inference hooks to give enterprises real-time DLP enforcement before Claude sees a prompt

· by Pondero Newsdesk

The short version

Anthropic shipped inference hooks on August 5, 2026, a beta Claude Enterprise feature that routes every employee prompt through a company-controlled security server for an allow-or-deny decision before inference starts.

Anthropic launches inference hooks to give enterprises real-time DLP enforcement before Claude sees a prompt

Anthropic on August 5, 2026 released inference hooks in beta for Claude Enterprise, a feature that routes every employee prompt through an organization's own security server for an allow-or-deny verdict before the model starts generating, and applies that same checkpoint to every tool-call response flowing through MCP connectors, skills, and plugins.

What

The mechanism sits at the infrastructure level, not the client. Per Anthropic's announcement, when a Claude Enterprise user sends a prompt, the request travels over a signed WebSocket connection to the organization's configured security server before it reaches the model. The server returns a verdict, and Claude only proceeds once it receives one. No client-side installation is required.

Tool calls get the same treatment. When Claude calls an external tool connected through MCP, a skill, or a plugin, the tool's response passes through the same inspection before the model can read it. That second checkpoint is new: prior to inference hooks, the contents of a tool response arriving from an external integration had no equivalent server-side inspection layer inside Claude Enterprise.

Coverage spans every Claude Enterprise surface from a single organization-level configuration. Chat, Claude Code, Claude Cowork, and other Claude Enterprise products all route through inference hooks once the feature is turned on. Per the announcement, earlier enforcement in Claude Code had been limited to client-side hooks, meaning the developer's local machine ran the check rather than a centrally managed server. Inference hooks closes that gap by moving enforcement server-side and extending it across all surfaces.

The webhook-based protocol is open with a documented schema, which means existing DLP vendors can build or publish integrations without requiring customers to abandon current security investments. Anthropic named Netskope, Palo Alto Networks, Proofpoint, and Zscaler as compatible platforms. Organizations can also point inference hooks at an in-house AI security server. Configuration options include shadow mode, which logs but always allows requests, enabling security teams to measure policy impact before turning on enforcement. Role-based exclusions, percentage-based rollouts, and adjustable failure-policy tolerances and timeouts are also available.

Inference hooks is available now in beta for Claude Enterprise customers. Anthropic published documentation at platform.claude.com/docs/en/manage-claude/inference-hooks.

Why it matters

The gap this closes was not theoretical. An enterprise running Claude Code, Claude Chat, and Claude Cowork had three separate enforcement surfaces before this release. Client-side hooks on Claude Code gave developers a way to run checks locally, but those hooks operated on the developer's machine, outside the security team's direct control. Chat and Cowork had no equivalent native enforcement. The result was that a compliance team could not make a consistent governance claim across all Claude Enterprise surfaces from a single policy server.

Inference hooks changes the answer to a question that comes up in almost every enterprise AI procurement review: can the organization's existing DLP stack cover this channel? Until August 5, the honest answer for Claude's full surface area was no. Now it is yes, provided the DLP vendor has built or will build a compatible integration.

The open webhook design is worth attention. The alternative path for Anthropic would have been building native integrations with each DLP vendor's proprietary APIs and shipping those as first-party features. That approach benefits customers who already use one of a short list of supported vendors, but it creates a dependency on Anthropic's integration roadmap and leaves everyone else waiting. An open protocol shifts that work to the DLP vendors themselves and to organizations with in-house security engineering capacity. Enterprise security buyers generally prefer that model, because it means the integration is not blocked by Anthropic's quarterly priorities.

The MCP coverage merits its own attention. Agentic Claude use cases increasingly involve Claude calling external tools to retrieve or write data. A customer database query result, an internal document retrieved by a connector, or a code repository read by Claude Code all arrive as tool responses. Before inference hooks, those responses moved from the external tool into the model without a centralized DLP checkpoint inside Claude Enterprise. That is now covered.

For operators evaluating whether to extend Claude to regulated business lines including legal, finance, or healthcare, the compliance calculation shifts. A DLP server already approved for email and SaaS traffic can now enforce the same approved policies on Claude traffic. That reduces the number of separate security approvals required to add Claude to a regulated workflow.

Context and reactions

Anthropic has been building toward a centralized enforcement layer for some time. Claude Code shipped client-side hooks earlier as a developer-facing tool for scripting pre- and post-response actions. Those hooks served a useful purpose for individual developer workflows but were not designed for centralized compliance governance. Inference hooks is the server-side, compliance-grade version of that concept, extended to every Claude Enterprise surface.

The broader market context is that enterprise AI governance has become a distinct buying category. Microsoft's Copilot for Microsoft 365 has data boundary controls, and Google Workspace's Gemini integration has its own policy enforcement layers. Anthropic's approach with inference hooks differs in one concrete way: the enforcement point is vendor-agnostic on the security side. An organization that uses Netskope for SaaS DLP and Palo Alto for network security can apply both to Claude traffic through the same webhook server, rather than choosing a Anthropic-native policy tool.

Token Security, an AI access security firm, published an analysis noting that the webhook-based design allows security vendors to build integrations without requiring customers to change their DLP stack, per their post. Anthropic did not announce pricing changes for inference hooks. The feature ships as part of Claude Enterprise rather than a separately priced add-on, per the announcement.

What to watch next

The beta label means the API surface may change before general availability. Two specific milestones are worth tracking.

First, published latency data. Inference hooks adds a webhook round-trip to every prompt before inference starts. In low-latency agentic workflows where Claude calls tools in rapid succession, each added round-trip has a compound cost. Anthropic has not published latency benchmarks showing the practical impact. Those numbers will shape whether inference hooks is viable for latency-sensitive use cases or primarily suited to lower-frequency compliance reviews.

Second, confirmed DLP vendor integrations. Netskope, Palo Alto Networks, Proofpoint, and Zscaler were named in the announcement, but Anthropic did not publish integration-availability dates for any of them. Security teams will need vendor-confirmed integrations, not just named compatibility, before they can complete their internal procurement reviews. The first confirmed integration timeline from any named vendor marks the next concrete milestone for enterprise buyers evaluating whether to route their Claude Enterprise traffic through inference hooks now or wait.

Sources