Anthropic Launches User Profiles in the Claude API So Safety Enforcement Can Target One End Customer Instead of a Whole Organization

Overview

Anthropic has launched user profiles for the Claude API, according to its October 20, 2026 platform release notes. The release notes say the feature lets developers “tell Anthropic which of your end customers each API request is for,” so that a safety enforcement action can apply to one end customer instead of an entire organization. The same entry also covers a beta set of organization webhooks and notes that inference hooks are out of beta for Claude Enterprise organizations.

The problem profiles address

Per the user profiles documentation, platforms, gateways and resellers send requests for their end customers under their own organization. The documentation states that “Anthropic runs automated safety classifiers over API traffic,” and that without a profile on the request, one end customer can trigger an enforcement action that applies to the whole organization. With a profile attached, an action such as a restriction or a ban can instead apply to one end customer, leaving the rest of the organization’s traffic unaffected, according to the same page.

How it works

According to the documentation, a developer registers each end customer once with POST /v1/user_profiles, stores the uprof_ ID that comes back, and sends that ID in the anthropic-user-profile-id header on every request made for that customer. The SDKs and the ant CLI support the new endpoints, per the release notes. The documentation says the header also works on token counting, Message Batches and OpenAI SDK compatibility requests, and that on a batch it applies to every request in the batch.

The documentation defines two profile types through an access_type field:

• application, the default, is for a signed-up user of the developer’s product, with one profile per user.
• passthrough is for a company to which the developer resells API access, with one profile per company rather than per person. It is not the default, so it has to be sent explicitly.

The documentation says every profile field is optional and that a profile does not contain conversation content. It also says Anthropic does not verify the details a developer sends. Developers who cannot share a name or email address can send hashes instead, and the page says a profile has no field for the email address itself. The page adds that profiles are not billable and do not affect pricing.

The documentation recommends keeping end-customer traffic in one organization and a developer’s own internal usage in another, because profiles belong to an organization rather than a workspace. For a request that comes from one identifiable person inside a passthrough company, it advises also sending an identifier in the Messages request’s metadata.user_id field.

Platform differences

The documentation says profiles can be created on the Claude API, Claude Platform on AWS, and Microsoft Foundry deployments hosted on Anthropic. The Claude API, Claude Platform on AWS and those Foundry deployments check the ID on every request and return a 400 invalid_request_error if it is malformed or belongs to a different organization.

Amazon Bedrock and Google Cloud do not have the /v1/user_profiles endpoints, according to the same page, so profiles for those platforms must be created on the Claude API or Claude Platform on AWS. On Claude in Amazon Bedrock the ID goes in a user_profile_id field in the request body. The documentation says Bedrock and Google Cloud accept any well-formed ID, and that a request with no ID “goes through unattributed,” with any resulting enforcement action applying to the whole organization.

Deleting a profile permanently removes what the developer supplied, but the documentation says it does not remove the trust and safety records Anthropic keeps about that profile ID. It also says inference requests carrying a deleted ID can still succeed for up to about 10 minutes.

Organization webhooks

The organization webhooks documentation describes two event types, user_profile.enforcement_action_opened and user_profile.enforcement_action_resolved. The release notes say the endpoints are in beta under the webhooks-2026-10-07 beta header. The documentation carries a note that reads: “Event delivery isn’t live: Anthropic doesn’t send the two events, so an endpoint you register receives nothing.” Developers can register endpoints and build handlers from the documented event format in the meantime.

Per the documentation, events carry IDs rather than the action itself: data.id is the profile ID and data.enforcement_action_id starts with enac_. A handler is expected to retrieve the profile to see the action’s type, either a restriction or a ban. Deliveries follow the Standard Webhooks specification, must be answered with a 2xx status within 5 seconds, and are retried a few times over the next few minutes before being dropped. An organization can have up to 32 endpoints, and access requires the feature to be turned on for the organization, which is requested through an Anthropic account team.

What We Don’t Know

The cited pages do not say how many organizations have enforcement actions switched on, and they do not describe the criteria Anthropic’s classifiers use. They also do not give a date for when webhook event delivery will go live.

Related coverage

The release notes’ note on inference hooks follows Anthropic’s earlier launch of the feature in beta, which let enterprises screen Claude prompts before they reach the model.