Cloudflare has put an identity-aware version of its AI Gateway into open beta and made a companion feature, User Insights, generally available to every AI Gateway customer at no extra cost, the company said in a blog post during its Agents Week.
AI Gateway sits between applications and model providers. The new integration lets a customer put a custom domain in front of the gateway and protect it with Cloudflare Access. Every authenticated request then carries the user’s verified Access ID, which the gateway adds to request metadata as cf.user_id so logs, analytics and spending can be filtered by the person who made the call.
That replaces shared API keys as the only record of who is using a model. “Shared API keys make it almost impossible to tell who is using an AI service or apply the access rules we already have for employees,” said Max Baumgarten, a staff security engineer at Flexport, an early adopter.
How it works
According to Cloudflare’s configuration guide, browsers attach the Access token as a cookie after login, while coding agents and other non-browser clients send it as a bearer token. The cf.user_id value is the JWT subject claim, not an email address. Service-token requests don’t get one because they don’t represent an individual user, and customer metadata keys starting with “cf.” are reserved and not saved. Traffic sent to the default gateway.ai.cloudflare.com endpoint isn’t protected by Access.
With identity attached, administrators can give each user a spend limit and either block requests or fall back to a cheaper model when it is reached. Cloudflare said group-based limits and model access tied to identity-provider groups are coming “in the near future.”
User Insights scores sessions rather than single requests. It compares each session with the account’s own 95th-percentile session cost over the previous 30 days and treats anything above twice that figure as a candidate anomaly, but only if the session also ranks in the most expensive 1% across the organization. A dollar floor keeps tiny accounts from setting off alerts over a few cents.
Cloudflare said it is also building task-based routing that sends each request to the cheapest model that can handle it, plus prompt classification that tags traffic as coding, writing or other categories. It recommends starting in monitoring mode to learn baselines before enforcing anything.
One stat the company cited for why it matters: a Stanford report found 59% of organizations named knowledge gaps as their biggest obstacle to responsible AI governance.
WebProNews is an iEntry Publication