Catching rogue AI behavior with identity-aware analytics
15:00 · August 5, 2026 · Cloudflare AI Blog
Identity-aware AI Gateway is now in open beta. User Insights turns that traffic into a behavioral baseline for every person and agent, and flags insider risk the moment it appears.
Summary
Cloudflare has introduced two capabilities that address the difficulty of spotting unusual AI activity in large organizations. Its Identity-aware AI Gateway, now in open beta, combines with the generally available User Insights feature to create per-identity behavioral baselines from traffic already passing through the gateway. By routing requests to models from OpenAI, Anthropic, Google and Workers AI through a single control plane, the system records who or what is making each call and compares current patterns against each account’s recent history.
Identity comes from Cloudflare Access, which attaches a verified user identifier to every request. This allows administrators to apply per-user spend limits, filter logs and analytics by individual, and enforce existing identity-provider policies without deploying additional infrastructure. Early adopter Flexport noted that shared API keys previously made it nearly impossible to attribute usage or apply access rules; the Access integration replaces that anonymity with authenticated identities that map directly to existing employee groups.
User Insights turns the same traffic into an anomaly-detection feed. It scores sessions rather than isolated requests, using each account’s 95th-percentile session cost over the preceding 30 days as its personal baseline. Deviations above twice that threshold, when they also exceed a minimum dollar value, surface as candidates for review. The approach distinguishes between the tight, regular patterns typical of automated agents and the more variable behavior of human users, reducing false positives while highlighting potential rogue agents, credential misuse or insider-risk scenarios.
The resulting view filters out routine activity and presents only accounts that have departed from their own norms, giving security and finance teams a focused starting point for investigation. Additional planned features, such as task-based model routing and prompt classification, aim to extend the same data stream toward cost optimization and clearer separation of intended versus unintended use.
Why it matters
Directly actionable for Dutch security teams managing AI spend, governance, and insider threats; supports EU-aligned responsible AI practices via identity and anomaly detection on existing traffic.
