Bottom line: With Inference Hooks, Anthropic enables centralized, webhook-based real-time inspection of prompts and tool responses across all Claude Enterprise surfaces, integrable with existing DLP systems such as Netskope, Palo Alto Networks, Proofpoint or Zscaler.
Anthropic is giving Claude Enterprise customers a new control layer called “Inference Hooks,” allowing compliance and security teams to inspect and block every prompt and every tool response before it is processed by Claude. The feature closes a gap, since native inline controls previously existed only for Claude Code via client-side hooks.
According to Anthropic, Inference Hooks route every inference request through a signed WebSocket connection to a security server controlled by the company. Before Claude begins generating a response, the system sends the prompt and context to this server; the server returns a verdict — “allow” or “deny” — which Claude enforces in real time. The same inspection mechanism also applies to tool calls: responses from tools connected via MCP connectors, skills, or plugins are checked before being passed back to the model. According to Anthropic, the feature covers all Claude Enterprise surfaces, including Chat, Claude Code, and Claude Cowork, and is enabled centrally at the organization level rather than configured per product.
What matters most to CISOs is the integration capability: Inference Hooks is based on an open, webhook-based protocol with a published schema. This allows the feature to be connected to existing DLP infrastructure, such as Netskope, Palo Alto Networks, Proofpoint, or Zscaler, or to custom-built AI security servers. Previously, security teams had to implement separate control mechanisms for each Claude surface or run agents per product; with a central configuration, this extra effort is eliminated, and existing DLP programs can be extended directly to Claude usage, preventing sensitive data from leaking through uncontrolled channels.
For rollout, Anthropic provides several control mechanisms: a shadow mode that initially lets all requests through and only logs them, role-based exceptions, and percentage-based rollouts for gradual introduction. In addition, failure-policy tolerances and timeouts can be adjusted to match the organization’s respective risk profile. Inference Hooks is available immediately as a beta version for Claude Enterprise customers; the DLP server is configured using documentation provided by Anthropic. Security vendors can develop their own integrations based on the documented schema, allowing customers to connect their organization directly to the respective platform.
Source: claude.com · Published August 4, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.