Cloudflare announced that Identity-aware AI Gateway is available in an open beta through integration with Cloudflare Access, while the User Insights feature is generally available to all AI Gateway customers at no additional cost. The two features use request traffic passing through the gateway to build a behavioral profile for each person or software agent, then draw administrators’ attention to accounts that deviate from their usual patterns.
Cloudflare says that knowing who is using AI and what they are doing represents an organizational and security challenge, citing a Stanford University report that found 59% of organizations considered knowledge gaps the biggest obstacle to responsible AI governance. The company believes addressing the problem requires a verified identity associated with every request, along with a baseline showing what normal usage looks like for each identity.
Verified Identity for Every Request
AI Gateway operates as a central control point for AI usage, with applications and developer tools sending requests through it instead of connecting directly to OpenAI, Anthropic, Google, and Workers AI models. Tools such as Claude Code, Codex, and GitHub Copilot can also be routed through the gateway, placing them under the same visibility and controls.
With Cloudflare Access integrated, a customer can place a custom domain in front of the gateway and protect it as they would any other application. This enables:
- Authentication through an identity provider that supports SAML, such as Okta or Entra, instead of creating and distributing Cloudflare API keys.
- Setting gateway access policies based on authorized users.
- Using a clean hostname such as ai.example.com without including the account or gateway identifier in the URL.
- Adding the authenticated Access user ID to the request metadata under the cf.user_id field, allowing logs, analytics, and spending to be sorted by the person who submitted the request.
Combined with spending limits, this can create a separate budget for each user, then stop requests or route them to a less expensive model when the limit is reached. Cloudflare says identity-provider groups will eventually be able to determine spending limits or the models available to each group, such as giving a machine-learning team access to advanced models or imposing a spending cap on a support team.
Max Baumgarten, a senior security engineer at Flexport, said shared API keys make it difficult to know which user is using an AI service and to apply employee access rules. He added that placing Cloudflare Access in front of AI Gateway links every request to a verified identity and allows existing identity-based policies to be used without creating a separate authentication system for each client.
User Insights Builds a Baseline for Each Account
User Insights appears inside AI Gateway and reads existing traffic without requiring additional setup. The feature tracks cost and sources of waste, such as low cache-hit rates or large context windows, but it also focuses on a different question: Is the account behaving normally relative to its own history?
The feature treats people and agents as having different patterns. An agent may summarize tickets every three hours on a regular basis, while a person may send varied requests and conduct long sessions when handling complex problems. User Insights therefore compares a session with the account’s own history rather than using a single threshold for all users.
The monitoring mechanism begins by recording session costs over the past 30 days, then using the 95th percentile (p95) for the account’s session costs. A session exceeding twice this baseline is considered a strong candidate for anomalous behavior. At the same time, Cloudflare uses an account-level ceiling at the 99th percentile (p99), with a minimum dollar threshold, so that a very small increase from a low-spending user does not trigger an alert.
Detecting Deviations Without Making an Automatic Decision
User Insights groups accounts that have broken their usual pattern into an anomalous-behavior digest, excluding increases that appear normal for a heavy user or have low financial value. A deviation could indicate that a service account has begun running more expensive sessions, or that a user’s spending has risen sharply and remained elevated for several days.
The feature does not determine user intent or block users automatically; instead, it presents unusual accounts to the administrator for investigation. The result could be a security incident or simply guidance toward a more efficient approach, such as avoiding sending an entire code repository when a snippet would be sufficient. Cloudflare emphasizes that spending and deviation visibility works without Cloudflare Access, but linking a request to a real identity turns an anonymous account identifier into a name that action can be taken against.
Availability and Next Steps
User Insights is available today in the dashboard to every AI Gateway customer at no additional cost, provided traffic passes through the gateway. Identity-aware AI Gateway with Cloudflare Access is currently available in an open beta. Cloudflare recommends starting in monitoring mode to learn the baselines before enforcing policies or usage limits.
The company is also developing intelligent routing based on task type, which selects the least expensive model that achieves the required result, as well as classifying requests into categories such as programming and writing. It says these classifications could help organizations understand the purpose of AI use and distinguish between commercial and personal usage, but intelligent routing and classification remain under development and have not been announced as available features in the current release.