Anthropic has put Inference hooks into beta for Claude Enterprise organizations, adding a server-side policy checkpoint before governed prompts reach the model.
The August 5 Claude Platform release notes say each governed prompt across claude.ai, Cowork, and Claude Code can be held for an allow-or-deny verdict from an organization’s AI security server before inference proceeds. Requests are signed, failure handling is configurable, and every denial is recorded in the compliance Activity Feed.
The docs describe the flow plainly. A user submits a prompt. Anthropic sends an HTTPS request to the configured security endpoint with the conversation transcript. The organization’s server returns an allow or deny verdict within the configured timeout, which defaults to five seconds. On deny, the request never reaches Claude.
Inline control changes the governance moment
Most enterprise AI controls are reactive. They log what happened, review transcripts after use, or scan output once sensitive data has already crossed a boundary.
Inference hooks move one decision earlier. The organization can block a governed prompt before the model processes it, using its own policy server or a vendor-operated AI security server.
That is especially relevant for Claude Code and tool-enabled agent work. The question is not only whether an employee pasted a secret. It is whether a session contains source code, customer data, unreleased financial information, regulated records, tool results, or context from attachments that should not enter a model workflow.
The beta has real limits
Anthropic’s docs say the security server receives transcript text, tool calls and results, and text extracted from attachments. It does not receive raw file bytes, image bytes, system prompts, or Anthropic-internal context.
The current hook event is prompt-side. Response-side enforcement is planned as a later event. That means Inference hooks can stop some sensitive inputs before inference, but they are not a complete output DLP system, hallucination monitor, or model-behavior judge.
The configuration page also matters. Organizations can use off, shadow, and enforcing modes. Shadow mode gives security teams a path to observe verdict behavior before blocking employees, which is usually the difference between a control that survives rollout and one that gets turned off after false positives.
