Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Claude goes through security first: Anthropic adds a Security Center to Enterprise as inference hooks go GA

On November 5 Anthropic gave Claude Enterprise a full security check-up: a Security Center that grades settings against its own recommendations, inference hooks that run prompts and tool calls past your own security tools first, centralized management for multi-organization identity, and EDR sensors inside the Cowork VM. The second half of the model race is about who can get inside companies.

Anthropic's official Security Center screenshot: 13 of 16 controls passed (81%), broken down by chat, Cowork and Claude Code.

On November 5, Anthropic gave Claude Enterprise a full security check-up: a security center (beta), inference hooks (generally available), centralized management (beta), one-click member sign-out, a self-serve IP allowlist, desktop controls, and a way for a company's own EDR sensor to run inside the Cowork virtual machine. This is not a routine feature update. It answers the hardest line on an enterprise procurement form: how do you prove the AI will not walk out with company data? Over the past year Claude Code has been writing code and Cowork has been running local tasks, and agents have become capable enough that the buyer has changed — from front-line developers to security teams. This batch of updates is aimed squarely at the latter.

Anthropic's official Security Center screenshot: 13 of 16 controls passed (81%), broken down by chat, Cowork and Claude Code.

1. The security center: turning "is it configured right?" into a table

The security center is a new page in Organization settings, and what it does is concrete: it gathers Claude's security settings for chat, Cowork, and Claude Code onto one page and checks each against Anthropic's recommended configuration. Every control lists its current value, the recommended value where there is one, and a link to the setting where you change it. Settings that are new or recently changed are flagged. The page re-checks the configuration every time it loads, so automatically checked controls reflect the current state; organizations on the Compliance API can read the same results through it.

The numbers in the official screenshot show what the page looks like: 13 of 16 controls passed, an 81% pass rate, with 10 of 13 passed at the organization level, 2 of 2 in Cowork, and 1 of 1 in Claude Code. The screenshot also shows 1,284 assigned users, 1,032 active users, 18 active connectors, and 4 organization owners — a dashboard for the person responsible for security, not for developers.

Beyond looking, the security center offers a way to review together. Attestations give admins a shared workspace for security reviews: work through the process once, recording the current state of each control, accepting risks, and flagging issues in one place. The whole page can be downloaded as a report for auditors and stakeholders, marking which controls the page checked automatically and which an Owner confirmed. That distinction matters: an automated check can only prove "the configuration looks like this"; a human confirmation means "someone takes responsibility for it".

But Anthropic states the limit up front: the security center is guidance based on its recommendations, not a certification, and not a guarantee of compliance with any law or standard. That sentence deserves to stand on its own — the table can help you find where your setup differs from the recommendation, but it cannot prove compliance to a regulator. Used as a health report it is exactly right; used as a certificate it is a misreading. Anyone who knows enterprise security can see the judgment in that line, and it is precisely where Anthropic is being honest: it did not dress up a comparison table as a compliance stamp.

2. Inference hooks: the only part of the batch that can actually block something

The security center can only see; inference hooks can stop things, and that is the essential difference. Hooks connect Claude to the data security tools an organization already uses, so data loss prevention (DLP) and security policies apply before prompts and tool calls take effect. There are two kinds: prompt hooks send a prompt to the organization's security tool for a check before the model starts generating, and pre-tool-call hooks check a tool call before it runs. The check covers not just chat but tool calls made through MCP, skills, and plugins — a tool's response is also checked before it goes back to the model.

Technically, the mechanism runs over a signed WebSocket connection: once an organization turns it on, every inference request sends the prompt and its surrounding context to the company's own security server, which returns one of two verdicts — allow or deny — and Claude only proceeds once it has one. Companies do not have to change tools; they point it at Netskope, Palo Alto Networks, Proofpoint, Zscaler, or a server they built in-house, and one configuration covers chat, Claude Code, Cowork, and other Enterprise surfaces. Rollout is gradual by design: shadow mode (watch without blocking), role-based exclusions, percentage-based rollouts, and customizable timeouts and failure policies.

How teams actually adopt it is worth a paragraph of its own. In Anthropic's article on hooks, Andrew Grimmett, Vice President of Information Security at Bandwidth, is quoted saying that inference hooks add a checkpoint to inspect what is flowing to Claude in real time, before the model ever sees it, letting the company move faster on AI without giving up control. The existence of shadow mode shows Anthropic understands a security team's first instinct is not "block immediately" but "show me what would get blocked" — run it in observe-only mode, learn the false-positive rate, and only then tighten. Role exclusions and percentage rollouts follow the same logic: a small group and a small share of traffic first, confirm normal work is not interrupted, then expand. Whether a security mechanism survives inside a large company often comes down to whether it offers this kind of "soft first, hard later" ramp.

One timeline detail deserves care: hooks were not born on November 5. Anthropic first released them in beta on August 5, and according to AI Weekly they became generally available on October 20; November 5 was the announcement packaged with the security center. Also according to AI Weekly, hooks carry a default timeout of five seconds, return only allow-or-deny verdicts, and do not inspect image-only content — those operational details come from secondary reporting, and this article flags them as such.

To judge the value of hooks, look at the gap they close. Until now, enterprise control over Claude was mostly after the fact: read the logs, review the records, reconstruct what happened. Hooks move the checkpoint to before the fact: the prompt is checked before it reaches the model, the tool call before it runs. For DLP that is a qualitative change — where employees could previously paste customer data into any chat box, there is now at least one gate inside Claude that the company itself defines. The limits should be stated too: hooks inspect prompts and tool calls, not what the model is "thinking"; and according to VentureAtlas's reading, hooks do not currently cover deployments on Amazon Bedrock or Google Cloud — also secondary sourcing, and companies should verify it in the official documentation before buying.

3. Centralized management and access controls: no more running between organizations

For large companies, the most practical item in this batch may be centralized management. A group often runs several Claude organizations: one for engineering, one for sales, another for a subsidiary. Each organization's identity and access settings used to be managed separately; now the parent organization gets its own admin console at admin.claude.com, where single sign-on, directory sync (SCIM), domain verification, groups, and Compliance API keys are set once and apply to the organizations underneath. It also supports multiple identity providers — an acquired company can keep its existing login method or migrate one domain at a time instead of switching everything overnight.

The permission design keeps a careful balance. The parent organization has two roles, owner and admin; admins can manage everything except assigning owner and admin roles. But a parent role does not grant access inside an organization by default, and each organization keeps its own members, roles, and settings. One exception is worth noting: Compliance API keys created at the parent can read data across organizations. In other words, the power to manage identity is pushed down, while the audit view is pulled up — classic group-style governance.

The three smaller access controls each target a real scenario. First, Admin API member sign-out: one call signs a member out of Claude on every browser and device, and Claude Code has to sign in again. But it is not a ban: API keys, cloud sessions already running, and scheduled routines are not ended, and the member can sign in again afterwards — a "disconnect first, investigate second" emergency action, not a punishment. Second, the self-serve IP allowlist: enterprise admins can now set it up and manage it themselves instead of going through Anthropic, limiting access to most Claude products to approved IP addresses. Third, desktop controls: admins can manage the most-used Claude Desktop settings without a device-management push, block personal and other outside Claude accounts on managed computers to prevent shadow IT, and limit which workspace folders people can use with Claude.

Each item on its own sounds like routine IT work; together they point to one thing: Claude is changing from "a tool employees install themselves" into "a tool the company issues". For independent developers and small teams, none of this batch applies — it is all Enterprise-only. But it sets the direction of the tool you use: the more capable agents get, the tighter the controls become, and today's Enterprise checkpoint is very likely tomorrow's default for everyone.

4. EDR inside the virtual machine: even the agent's "locked room" has to be visible

When Cowork and unified Claude work locally, they run shell commands and code in an isolated virtual machine on the user's computer for additional protection. Security teams then asked a reasonable question: they can see activity everywhere else on the endpoint, so why is this one place a blind spot? The answer in this update is that admins can now run their organization's EDR sensor inside the Cowork virtual machine, and the sensor reports activity there to the same console the security team already monitors.

Three implementation details matter. First, each time the virtual machine starts, Claude verifies the sensor package's signature from the EDR vendor before installing it — the sensor itself has to pass a check. Second, sensors run in detection mode: they report activity and do not block it. Third, this covers the virtual machine on the user's computer, not Cowork sessions in the cloud; according to Datastudios, seven vendors are supported, including CrowdStrike, Palo Alto Networks, SentinelOne, and Tanium, with the full list in Anthropic's Help Center.

"Detect only" is honest, and it is also a concession. A command an agent runs inside the virtual machine, blocked in error, would break the workflow in progress, so starting with observe-only is the reasonable first step. But "visible" is itself progress: until now, what an agent did inside the virtual machine was invisible to the security team; now at least the activity shows up on the screen they already watch. For a company, that is what makes a security team say yes — not a guarantee that nothing will go wrong, but a record to consult when something does.

5. Beta versus generally available: most of this batch is still a trial

Count the statuses: the only generally available item is inference hooks. The security center is in beta, centralized management is in beta, member sign-out is in beta, the self-serve IP allowlist is in beta, desktop settings management is in beta, and EDR in Cowork is in beta. That distribution is itself worth reading: Anthropic took the most mature piece — hooks, in beta since August and refined for three months — and made it generally available first, leaving the rest in trial.

For a company, beta means two things. First, features may still change, and problems should not be expected to be covered to the standard of a generally available service — and the security center itself says it is guidance, not certification. Second, centralized management setup cannot be undone, and the official text specifically warns admins to identify their parent organization admins before they start. Taken together, Anthropic's stance is clear: the mechanisms are yours to use, but the decision — and the responsibility — to put them into production is yours. A security team that writes beta features into a compliance commitment is handing in a trial product as a final answer, and that is a distinction it has to hold itself.

One more detail hides in the EDR section: sensors only detect, they do not block. On the surface that looks incomplete; underneath it is Anthropic conceding something about how agents work — block one command in error and a whole workflow breaks. Detect first is an admission that "we are not yet willing to decide for you what should be blocked", and it is more honest than a blocking interceptor with a high false-positive rate. The next things to watch are whether EDR moves from detection to blocking, and when the security center drops the beta label — that day will be when this mechanism truly graduates.

6. Why now: two things arrived together

The first is that agents are moving into the core of company workflows. Claude Code writes code and Cowork runs local tasks; agents are no longer experiments but part of production. The security team's question changed with that: from "can we use it?" to "who is responsible when something goes wrong, where does the data go, and who can prove it?" This batch reads almost line by line from a security review checklist: a configuration comparison table, a pre-flight block, unified identity, emergency sign-out, endpoint visibility — each item matches one row of the checklist.

The second is a coincidence of timing. According to AI Weekly, the day before the announcement, a report alleged that Claude Code was leaking user email addresses — through the User-Agent header. That claim is unconfirmed by Anthropic, and this article cites it only as secondary reporting; but the timing itself is worth noting: Anthropic is repairing trust on one side while selling control on the other. For a company in the middle of an IPO push, enterprise trust feeds directly into the valuation story — this site already covered Anthropic's pre-Thanksgiving IPO at a $1.25–1.75 trillion valuation on November 7, and this security update can be read as the other half of the same story: alongside growth numbers, the roadshow also needs evidence that large companies can use the product with confidence.

Put the two together and the conclusion is clear: the first half of the model race was about who is smarter; the second half is about who can get inside companies. Smartness answers "do people want to use it"; control answers "does the whole company dare to use it?" Anthropic did not release a smarter model this time — it released a full set of mechanisms for getting security teams to say yes. That choice is itself its judgment about the second half.

7. What this means for vibe coding

If you are an independent developer, none of this batch touches you in the short term, but it signals the direction. Today's Enterprise checkpoint is very likely tomorrow's default for everyone: prompts checked first, tool calls approved first, sensitive actions logged. Understanding that logic now beats being stopped by it later and having to read the documentation then.

If you use Claude Code inside a large company, this update means the days of "installing Claude Code quietly on your own" are ending. IP allowlists, desktop account controls, and EDR in the virtual machine all push a "personal tool" toward becoming a "company asset". The compliant path is the company's unified procurement and configuration, not working around the controls — people who work around them tend to show up by name in the audit report.

Three things are worth watching next, and all three are concrete. First, when the security center and centralized management drop the beta label — that will be the moment Anthropic formally hands in its homework to enterprise customers. Second, whether inference hooks grow from "connect your own DLP" into "we include a DLP for you", which decides whether smaller companies can afford it. Third, whether EDR moves from detection to blocking — once that step is taken, the tug-of-war between agent autonomy and security control truly begins. In one sentence: what Anthropic sold this time is not a smarter model but the ability to get a security team to say yes. The security center makes things visible, inference hooks guard the door, and centralized management and EDR make the whole thing runnable inside a large company. Whether models can get smarter can wait for the next round; whether they can get into companies, and stay there, this update has already answered.

Views 6Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality