Unveiling good and bad behaviors on the Agentic Internet (opens in new tab)
The Internet contains a mix of human, automated, and hybrid traffic, so classifying visitors simply as “good humans” or “bad bots” no longer works. Cloudflare argues that website owners should evaluate continuous behavior and build trust over time, rather than rely only on isolated risk checks. Its strategy combines transparent bot classification through BotBase with behavioral detection through Precursor. ## Risk and Trust Are Different - **Risk** measures how likely a request or action is to be harmful, often based on a momentary event. - **Trust** develops over time through reputation and observed behavior. - A suspicious action should not automatically lead to rejection if the actor has an established, trustworthy identity. - Cloudflare’s goal is to create incentives for safer behavior, from blocking malicious activity to enabling beneficial automated participation. ## Verified Bots and Transparency - Cloudflare defines verified bots and agents through two requirements: - They honestly declare who they are. - They do not abuse the trust they receive. - Transparency helps site owners decide which behaviors and data uses to permit. - BotBase is designed as a directory of all known bots and agents, not just approved ones. - Actors that violate expectations can lose their verified status and become harder to trust across Cloudflare’s network. ## Continuous Detection with Precursor - Precursor is a client-side system that continuously analyzes behavior throughout a session. - Its JavaScript is injected at the CDN, so customers do not need to manually place or rerun detection code. - Unlike CAPTCHAs and other one-time challenges, Precursor evaluates behavior in context over multiple pages. - Continuous analysis makes it more expensive for bot developers to imitate human behavior across an entire session. ## What Cloudflare Has Observed - In a 24-hour snapshot, Precursor processed **206 million evaluation events across 73,438 zones**. - Suspicious behavior frequently begins in the middle of a session, meaning point-in-time checks can miss it. - A single session may shift between human and automated behavior, such as when a user hands part of a shopping journey to an agent. - Site owners therefore need classifications based on use case, purpose, and data usage rather than a simple human-versus-bot decision. ## Understanding Human Cursor Behavior - **Precursor Trace** is an interactive demonstration of part of Precursor’s detection process. - It examines cursor acceleration, corrections, rhythm, and movement texture. - The demo illustrates how behavioral signals can distinguish natural interaction from automated movement. ## Adaptive Intelligence - Cloudflare indicates that its bot detection engines will soon provide more adaptive outcomes for requests identified as automated. - The excerpt ends before detailing how this upcoming capability will work. Website owners should treat automation as a spectrum: verify transparent actors, monitor behavior continuously, and make decisions based on trust, intent, and use case rather than isolated signals.
