Claude Enterprise now blocks sensitive data in real time: what that choice reveals for your business

On August 5, 2026, Anthropic launched inference hooks in beta for organizations on Claude Enterprise. The principle: every prompt sent to Claude — on claude.ai, in Claude Code, or in Cowork — is first routed to the security server the company itself configured, which returns an allow-or-deny verdict in real time before the request reaches the model. In practice, an employee who pastes a client file excerpt or a payroll spreadsheet into Claude can have that request blocked before the data ever leaves the company, with a message explaining why.

What the feature does, precisely

• →The company’s security server receives the conversation transcript (text, tool calls, text extracted from attachments — never raw files or images) and returns an “allow” or “deny” verdict within a configurable timeout, 5 seconds by default.
• →Rollout can be gradual: a shadow mode that blocks nothing, a growing percentage of inspected requests, or exemptions for specific roles.
• →Coverage spans chat, Claude Code and Cowork, on web, desktop and CLI, plus responses from MCP connectors and plugins — but not API-only Platform organizations, nor Claude on Amazon Bedrock or Google Cloud.
• →Data-loss-prevention vendors — Netskope, Palo Alto Networks, Zscaler — already offer ready-made integrations.

A signal bigger than the feature itself

That the leading generative AI vendor built a dedicated blocking layer for the exact moment sensitive data gets typed in says something about how mature the underlying risk has become. A 2025 study by Mario Silic, Dario Silic and Kathrin Kind-Trüller, From Shadow IT to Shadow AI – Threats, Risks and Opportunities for Organizations, published in Strategic Change, combines a survey of 140 professionals with interviews of 10 executives: it documents that unauthorized AI use by employees — shadow AI — takes hold precisely in companies that offer no sanctioned alternative, with a data-leak risk the authors describe as structural rather than occasional. A second study, published in 2026 in Technological Forecasting and Social Change by Glorin Sebastian and titled Digital shadow AI risk theory (DART), proposes a six-dimension framework for this risk — including “unintentional disclosure” and the gradual erosion of internal governance as unsanctioned use becomes routine.

What it means for your SME

Inference hooks, for now, only apply to organizations on Claude Enterprise — a tier out of budget reach for most French SMEs. But the signal matters more than the feature: if even the vendor judges a real-time blocking layer necessary, data leakage through an AI tool is no longer a theoretical risk raised in a management meeting — it is a scenario large organizations are already budgeting for. An SME without the means for a dedicated security server can apply a lighter version of the same principle — less automated, but just as effective at its own scale — by controlling data upstream rather than reacting after the fact.

• →Map, without judgment, which free or public AI tools your teams already use and with what data — the same exercise Anthropic’s security server automates, just done by hand.
• →Write a simple blocklist: which data categories (contracts, payroll, client files, health data) must never be typed into an AI tool the company hasn’t approved.
• →When you have a custom agent built — an AI support desk, automated email sorting, an operations watchdog — require that its data access scope be defined at design time, not left open by default: the same upstream-blocking principle, applied to a tool you control.
• →If you are already a Claude Enterprise customer, test inference hooks in shadow mode before switching to blocking, to measure the real volume of affected requests without disrupting your teams.

Data leakage through a public AI tool is not tomorrow’s risk: the studies cited above show it is already widespread in organizations that have put nothing in place. That the market’s leading vendor answers with a technical blocking layer is confirmation, not news — and an SME that writes its internal policy this week doesn’t need to wait for an SME-priced version of this tool to start reducing the risk.