On November 5, Anthropic made inference hooks generally available in Claude Enterprise. Once configured, a security server run by the company or its vendor evaluates requests and returns an allow or deny decision. Checks can happen before Claude processes a prompt and before it runs tool calls covered by the configuration. Anthropic announcement.
Where policy enters the workflow
Imagine asking Cowork to use a connector to send a document summary. The rule for inspecting what goes into the model and the rule for allowing the tool action can apply at different points in the same workflow. The latter requires Validate tool calls. The documentation also describes exceptions, including certain built-in tool calls, and excludes deployments on Amazon Bedrock and Google Cloud. Inference hooks documentation.
There is an architectural consequence too. The security server may receive conversation text, tool results and text extracted from attachments. That gives the information the company wants to govern another destination. Teams operating the integration can map who can access that server and what it logs before extending enforcement across the organization.
What if the security server stops responding?
Administrators can continue without inspection or block the request when the server is unreachable or times out. The initial configuration allows requests through on failure and waits up to five seconds for a verdict. Shadow mode lets teams evaluate verdicts during rollout without blocking users. Configuration guide.
A pilot can answer a few concrete questions. How much time does the check add? Which legitimate requests would it block? Which workflows need to stop if inspection fails? The choice depends on the consequences of letting sensitive data through and the consequences of stopping work.
The package also includes a security center and a console for managing identity and access across organizations. Both are in beta. The security center compares settings with Anthropic's recommendations. To evaluate hooks in practice, take a representative request and see what happens when the policy allows it, when it blocks it and when the server stops responding.