Anthropic user profiles confine Claude bans to one customer

Anthropic says platforms and gateways that sell Claude access can now attribute each API request to an individual end customer, so that a safety restriction or ban lands on that customer rather than on the whole organization. The feature, called user profiles, appears in the Claude Platform release notes for 20 October 2026, with API version 2026-10-21 following on 21 October. The details below come from those notes and Anthropic's user-profiles documentation.

The problem it addresses

According to the documentation, requests a platform sends for its customers arrive under the platform's own organization. Anthropic runs automated safety classifiers over API traffic. Without a profile on the request, the docs say, one end customer can trigger an enforcement action that applies to the platform's whole organization. For a company running a customer-facing product on Claude, that is a shared-fate risk: one customer's misuse can affect everyone.

How it works

• Create one profile per end customer with POST /v1/user_profiles, ideally in the onboarding flow. The call is not idempotent, so repeating it creates a duplicate even if the reference_id matches.
• Send the returned uprof_ ID in the anthropic-user-profile-id header on every request made on that customer's behalf.
• Choose the profile type: application (default) for each signed-up user of your product, or passthrough for each company you resell raw API access to, one per company, not per person.
• Do not create a profile for your own usage. Anthropic counts a request with no profile as yours, and the docs advise keeping end-customer traffic and internal usage in separate organizations, because profiles belong to an organization, not a workspace.
• User profiles now carry enforcement_status and enforcement_actions fields, and the list endpoint can filter by status: unrestricted, pending, restricted or banned. Organization webhooks, in beta, fire when a profile is restricted or banned and when that ends.

Limits to know

• Anthropic says profiles do not change billing or pricing, and contain only what you send plus program and enforcement status, not conversation content. It does not verify the details you send.
• Profile endpoints are available on the Claude API, Claude Platform on AWS, Microsoft Foundry hosted on Anthropic and Claude Models on Google Cloud (Anthropic managed). Amazon Bedrock and Google-managed deployments do not have them, so profiles must be created elsewhere.
• The webhooks are in beta and access is requested through an Anthropic account team.
• The docs say a ban has no end date and cannot be cleared by the end user. They do not say how often enforcement happens or how appeals work, so ask your account team.

For teams that proxy or resell Claude, the practical step is an audit: list which customer traffic reaches Claude without a profile ID, since that traffic is still exposed to organization-wide action.