Claude Enterprise can send each prompt to a company’s server before the model runs
Anthropic’s product article, dated 5 November, makes inference hooks generally available and puts a security center in beta. A prompt hook posts the transcript to an address the company chooses and waits, five seconds by default, for allow or deny. The docs say image-only attachments are not inspected, a verdict cannot redact a prompt, and the hooks do not run on Bedrock, Google Cloud, or the Claude Platform API.
Short2 minIn-depth
Priya Shah2 min read
- Anthropic
- Claude
- Enterprise
- Security

Photo: AnthropicSource page
Anthropic’s article on Claude Enterprise security controls is dated 5 November. The part that is generally available is inference hooks. A prompt hook sends a prompt from supported Claude products to a security tool the organization already uses, and a pre-tool-call hook checks a tool call before it runs. The security center, a page in organization settings that compares chat, Cowork and Claude Code settings with Anthropic’s recommended configuration, is in beta. The article says the page is guidance, “not a certification or a guarantee of compliance with any law or standard.” Organizations with the Compliance API can read the same results. An admin can download the page as a report. The report marks which controls the page checked and which an owner confirmed.
The docs are more precise about the hook. It is for Claude Enterprise only. Turning it on requires the organization:manage permission, which the docs say only the Owner and Primary owner hold. When someone submits a prompt, Anthropic sends an HTTPS POST to an address the organization configures. The body is the conversation transcript. Once the organization has generated a signing secret, the request is signed to the Standard Webhooks specification. The server has a timeout the organization sets, five seconds by default, to answer allow or deny. A denial never reaches the model. The user sees the reason the server put in deny_reason, then a standing message the admins wrote, or a default that tells them to contact an admin. The denial is written to the activity feed. If the server is down, slow, or returns an error, a setting chooses whether the request is blocked or allowed through without inspection. Repeated failures trip a circuit breaker that stops calling the server and applies that same failure setting until the server answers again.
What the server does not receive is the practical limit. Attachments arrive as metadata and extracted text, not as the file or the image, so a screenshot of a document is not inspected. A verdict is allow or deny. It cannot rewrite or redact the prompt. The hooks do not cover Claude Platform API organizations, and they are not available on Amazon Bedrock or Google Cloud. One hook is supposed to cover the Enterprise products, including claude.ai, Cowork, Claude Code, Claude for Microsoft 365, Claude Tag and Claude in Chrome. In voice mode, the docs say a small share of tool calls is not sent, and a denial usually ends the call with a general error rather than the reason. A blocked message stays in the chat and is sent again with later messages, so the person has to edit it out or start again. Some of Anthropic’s own calls — hosted Claude Security scans, Code Review on GitHub pull requests, and the model calls behind smart reports — are not governed requests. A session a person opens to fix a finding is governed. The article’s other beta controls sit beside the hooks: a parent-organization console at admin.claude.com, an Admin API call that signs a member out of browsers and devices without ending API keys or cloud sessions already running, a self-serve IP allowlist, and a block on personal Claude accounts in Claude Desktop on managed computers.


