The instructions you typed into Claude can now be pulled by every organisation you belong to
Anthropic’s Compliance API can now return the text each member of a Claude Enterprise organisation typed into Instructions for Claude in claude.ai, and because that setting belongs to the account rather than to the organisation, the same text comes back under every organisation the person belongs to. Anthropic’s platform release notes say the endpoint is in beta for Claude Enterprise organisations, in an entry dated October 15.
Anthropic lists the new endpoint as beta for Claude Enterprise organisations. Image: Anthropic Image: Anthropic
The call is GET /v1/compliance/organizations/{org_uuid}/user_preferences, and it needs an existing Compliance Access Key carrying the read:compliance_user_data scope. Two fields come back for each member: conversation_preferences, which is the claude.ai setting, and cowork_global_instructions, which is Cowork’s Global instructions.
Account-level means it crosses organisation lines
“Both are account-level settings, so the same text is returned under every organization a member belongs to,” Anthropic’s documentation says. A Claude Enterprise tenant can hold several linked organisations, so somebody who belongs to more than one of them wrote that line once for all of them.
The key is not something an individual member holds. Anthropic says primary owners and organisation owners create Compliance Access Keys in claude.ai, and that a primary owner’s key can cover every organisation under the parent organisation while an organisation owner’s key covers only their own.
Anthropic’s own warning about that scope is blunt. A key with read:compliance_user_data “can read every chat, file, project, and session transcript in every linked organization, including content the primary owner has not seen”, and Anthropic tells customers to treat such keys like production database credentials.
What comes back, and what Anthropic does not say
By default the response lists only members who have set at least one of the two values, and passing include_unset=true lists every current member with null where nothing is set. Up to 10 user IDs narrow a call through user_ids[], results paginate 20 at a time up to a maximum of 100, and Anthropic warns that a short or empty page can still not be the last.
The only published example of such a value is “Keep answers brief and cite sources.”, returned against a member’s email address. Nothing in the documentation says members are told the field can be read, and nothing establishes that an employer has called the endpoint.
It is a new read endpoint rather than a change to what Claude keeps, and it reaches Claude Enterprise organisations only. Somebody using a personal Free, Pro or Max account sits outside it.
Anthropic also steers customers away from the older route. It calls the audit log export that owners can download from claude.ai “significantly narrower than the Compliance API: a capped lookback window, CSV download only, and no access to chat, file, or project content”, and tells them to standardise on the API for ongoing use.
Every Claude surface keeps arriving at the same API
In September the Compliance API started returning transcripts of Claude in Chrome sessions, also in beta for Claude Enterprise, after Cowork and Claude Code sessions on users’ own machines arrived in August. Each of those is a record of work, logged as it happens.
This one is different in kind, because a member typed it. Instructions for Claude reads like a personal preference in claude.ai, and on an Enterprise account it is now a field with an endpoint and a scope.
Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.