August 5, 2026, (Inside AI) — Cloudflare is launching two new features to help enterprises detect and investigate anomalous AI usage: identity-aware AI Gateway with Cloudflare Access in open beta, and User Insights, now generally available at no extra cost. Together, they aim to solve a persistent blind spot in AI governance by tying every API request to a verified user identity and flagging behavioral deviations from that user’s own historical patterns.
The core problem is one of visibility. Organizations often rely on shared API keys, making it nearly impossible to attribute spend or unusual activity to a specific person or agent. A Stanford University report found that 59% of organizations cite knowledge gaps as their biggest obstacle to responsible AI governance. Without identity, a 10x usage spike is just a number on a bill, not a security incident or a misconfigured agent.
Cloudflare’s AI Gateway already acts as a central control plane, routing all model requests through a single point for logging, caching, and rate limiting. The new integration with Cloudflare Access adds a layer of authentication. Administrators can place a custom domain in front of the gateway, enforce SAML-based logins via providers like Okta or Entra, and set granular access policies. Every request then carries a verified user ID, which appears in logs as cf-access-user-id. This turns anonymous traffic into attributable actions.
“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, Staff Security Engineer at Flexport, an early adopter. “Putting Cloudflare Access in front of AI Gateway gives each request an authenticated identity and lets us use our existing identity policies at the gateway. Our teams can adopt AI tools without creating a separate authentication system for every client.”
With identity in place, User Insights builds a behavioral baseline for every account, whether it belongs to a human developer or an automated agent. It tracks session costs over a rolling 30-day window, calculating each account’s 95th percentile (p95) spend. Any session exceeding twice that personal baseline and also landing above the organization-wide 99th percentile is flagged as anomalous. This dual-threshold approach filters out noise: a $50 session from an agent that normally spends $5 is a 10x spike worth investigating, while a $500 session from a heavy user might be routine. Cloudflare’s internal analysis showed that the account-level p99 ($200) ensures alerts only fire for sessions that are both statistically unusual and financially material.
The system is designed to adapt. As a user’s habits change, their rolling p95 and 2x threshold move with them, avoiding static rules that generate false positives. A dollar floor also prevents micro-accounts from triggering alerts on tiny absolute changes. The result is a “rogue behavior feed” that surfaces only the accounts deviating from their own norms, without blocking any action. It’s a detective control, not a preventive one, leaving room for human judgment.
This approach reflects a broader industry shift toward behavioral analytics in AI security. Traditional policy-based tools catch known bad actions, but they miss the subtle signs of credential compromise or an agent going off the rails. By focusing on deviations from individual baselines, Cloudflare is targeting the gray zone where most insider threats and misconfigurations first appear. The methodology resembles user and entity behavior analytics (UEBA) in cybersecurity, but applied specifically to AI consumption patterns.
Cloudflare is also developing follow-on capabilities. Task-based smart routing will analyze incoming prompts and direct them to the most cost-effective model, potentially saving organizations money without sacrificing quality. Prompt classification will categorize requests into buckets like coding, writing, or summarization, adding context to anomalies and helping detect misuse, such as personal projects or data exfiltration disguised as legitimate AI usage. These features are in active development, with no release dates announced.
The identity-aware gateway is in open beta, and User Insights is available immediately to all AI Gateway customers. Cloudflare recommends starting in monitoring mode to establish baselines before enforcing spend limits or access rules. For organizations already routing traffic through AI Gateway, the new insights require no additional setup.
