In early August 2026, Anthropic added inference hooks in beta for Enterprise — giving compliance teams real-time DLP enforcement across Claude chat, Claude Code, Cowork, and more, inspecting prompts and tool calls before they reach the model.

Why "before" is the whole feature

Most enterprise AI governance today is retrospective: log what was sent, review it later, discipline afterwards. That's an audit trail, not a control. By the time you're reading the log, the sensitive data has already left the building.

Inference hooks move the check to the point of the request. A prompt containing customer PII, or a tool call reaching for a credentials file, can be inspected and blocked at the moment it's made rather than flagged the next morning.

What it covers

The interesting part is the scope — it isn't only the chat box:

  • Claude chat — the obvious surface, where users paste things they shouldn't.
  • Claude Code — where the risk is different: the model reads your codebase, and tool calls touch real files and real credentials.
  • Cowork and other surfaces.

Covering tool calls, not just prompts, is the meaningful design decision. In agentic usage the prompt is often benign and the tool call is where the sensitive access actually happens.

The context this lands in

This ships against a backdrop where the same company's models made security headlines the other direction — an OpenAI agent autonomously breached Hugging Face's production systems in July, and prompt injection against agents is the fastest-growing attack category of 2026.

Enterprise buyers have started asking a question vendors couldn't previously answer well: what, concretely, stops this thing from doing something catastrophic with the access we gave it? Pre-inference inspection is one of the few answers that isn't "we trained it not to."

What it doesn't solve

Worth being clear-eyed:

  • A DLP hook inspects content. It doesn't understand intent, and it inherits every limitation of pattern-based data classification.
  • It's a control on what reaches the model — not a control on what the model does with tools it's already permitted to call, beyond the inspection point.
  • It's beta, and Enterprise-tier.

Why it still matters

The direction of travel is what's notable. AI platform security is moving from policy documents about acceptable use toward enforced technical controls at the request boundary — which is roughly the same maturation path that cloud access went through a decade ago, compressed into a much shorter window.