The News, Explained
On November 5, Anthropic announced new security features for Claude Enterprise. The security center, now in beta, brings company settings together for review. Inference hooks, now generally available on Enterprise, let a company apply its own security rules before Claude generates a response or runs a tool. Source
The security center compares current settings for chat, Cowork, and Claude Code with Anthropic’s recommendations. The page is read only: it shows how the settings compare and where to change them. For a setting it cannot read directly, such as one set in a device management tool, an organization owner can record a manual check. The downloadable PDF separates automated checks from owner confirmations. It can support an audit, but does not confer a security certification. Source
An inference hook acts while Claude processes a request. A hook is a connection that asks a separate system for a decision before moving to the next step. For a request covered by the rules, Anthropic sends the conversation and text extracted from attachments to a server run by the company or its security vendor. An approval lets the request reach the model; a denial stops it first. Checking tools before they run requires a separate setting, and that stage can omit some built-in tools. Organizations using the Claude Platform API are outside the scope of inference hooks. Source
Imagine an employee pastes a confidential number into a chat. If the company’s rules detect and reject the number, the request stops before Claude starts answering. Put the same number solely in a screenshot, and this check will miss it because the server never receives the original image. If the server fails to respond, the company can configure Claude to block the request or let it continue without inspection. Source
OYOPICK’s Take
Meeting a recommended setting and stopping sensitive information in a live request produce different kinds of evidence. The first describes how Claude is configured; the second shows where a company’s rules actually work. A useful evaluation would try the same fictitious confidential number in plain text and in a screenshot, then repeat the exercise with the inspection server unavailable. A successful text test may still leave images needing separate protection. For an outage, the company has to decide when keeping work moving matters more than requiring an inspection. Source Source