Why these updates matter more than the models
In the same weeks that Fable 5.2, Sheets and Meetings came out, Anthropic released a set of controls for Claude Enterprise that made much less noise. For anyone bringing Claude into a structured company, they are probably more important.
The reason is simple. In the assessments we run with clients, the question that stalls projects is rarely which model is best. It's who controls what goes in and out, how I prove it to an auditor, how I attribute costs. Until recently many of these answers required custom work. Now part of it is inside the product.
There are five new things: security center, inference hooks, centralized management, access controls, project tags. Let's go through them one by one, with their limits, because in documentation limits are in small print and in security assessments they are the part that matters.
Security center: configuration compared with the recommended one
The security center, in beta, is a page in organization settings that compares the configuration of chat, Cowork and Claude Code with the one Anthropic recommends.
For each control it shows the current value, the recommended one and a link to change it. It highlights new or recently changed settings, which is useful given how often features arrive switched on by default. Admins can record the status of each item, explicitly accept a risk and flag issues in a shared area.
The most useful part for anyone facing audits is the downloadable report, which separates controls verified automatically from those confirmed manually by an owner. It can also be read through the Compliance API.
Anthropic notes it is not a certification or a guarantee of compliance with laws or standards. It is a tool to know how the system is configured and document it, and that alone saves days of work.
Inference hooks: checking data before it reaches Claude
Inference hooks, now available to all Enterprise customers, are the most important technical addition.
There are two kinds. Prompt hooks send each request to a company security tool, for example a data loss prevention system, before it reaches the model. Pre-tool-call hooks check the actions Claude is about to take with connected tools before they run. A single hook covers claude.ai, Cowork, Claude Code, Claude for Microsoft 365, Claude in Chrome and Claude Tag.
In practice: if someone pastes a document into Claude with data that shouldn't leave, the rule the company has already defined in its DLP system can block it.
Two limits to know. Hooks don't apply to organizations using Claude through Amazon Bedrock or Google Cloud, nor to API usage of the Claude Platform. And according to an independent analysis, if the external tool doesn't respond the request goes through instead of being blocked. The official documentation doesn't clarify the point: check it before presenting hooks to an auditor as a blocking control.
Need to get Claude past your security team?
30 minutes to discuss your specific case.
Multi-organization groups, access and EDR
For groups with several companies or several separate Claude organizations there is centralized management, in beta. A parent organization gets its own console at admin.claude.com from which it manages single sign-on, user provisioning, domain verification, groups and Compliance API keys for all linked organizations. Each organization keeps its own members and controls, and the group admin role does not grant automatic access to content. Note: the linking procedure cannot be reversed.
On access controls, also in beta, there are four practical additions. You can sign a user out of every browser and device with an API call, useful when a laptop is lost. Admins manage the IP allowlist themselves. You can block personal Claude accounts in the desktop app on company computers. And you can restrict the folders Cowork works on.
Finally, the company's EDR sensor can run inside the Cowork virtual machine on the user's computer, in detection mode: it flags, it doesn't block. This also applies to Team plans.
Project tags: spend by cost center or client matter
On November 6 a small feature arrived that is very concrete for professional firms and consultancies: project tags, in beta on Enterprise.
You attach a label to a Claude project, for example cost center or matter number, with its value, and the spend for all work done in that project is attributed to that code. A project can carry up to 16 tags, but each tag holds one value. Reports are in settings, under spend by project, and also available via API.
The limits are clear. It only applies to work inside a project: a chat outside projects isn't attributed. Claude for Office and Claude Code on the web are excluded. You can't split a project's spend across two matters. And values aren't checked against the accounting system, so a wrong code stays wrong.
For a law firm that wants to pass AI costs on to the client, or a company that wants to know how much each business unit spends, it is still the first native tool to do it.
What I'd turn on first
If you run a Claude Enterprise organization, this is the order I'd suggest.
First the security center: it takes ten minutes, tells you where your configuration differs from the recommended one and gives you a document to attach to internal reviews.
Then inference hooks, if the company already has a DLP tool. Connecting them is the fastest way to apply to Claude the same rules that cover email and cloud. First, though, clarify with the vendor what happens if the tool doesn't respond.
Then project tags, if management control asks to attribute costs. It's worth introducing them early, before habits of working outside projects set in.
I'd evaluate centralized management calmly, precisely because there's no going back. On Enterprise deployment in general there is the guide from people who do it every day, and on costs the comparison of Pro, Team and Enterprise.