Your Cursor Movements Might Give You Away! Cloudflare's New Bot Detector "Precursor" Goes Trust-Based

Hey there, it's Shii! Today I've got a pretty serious announcement from Cloudflare about how it tells bots and AI agents apart on the web. Did you know your cursor movements are being watched, too?

Cloudflare Blog blog.cloudflare.com

What was announced?

What Cloudflare posted on its Blog is a big shift in how it judges bot and AI agent traffic: moving away from a one-shot "Risk" assessment and toward continuously evaluating "Trust" throughout an entire session.

At the center of this shift are two things: "BotBase," a comprehensive ledger that tracks every known bot and agent, and "Precursor," a detection engine that keeps watching behavior for the whole session. On top of that, Cloudflare also released a simulation tool called "Precursor Trace" that lets you check for yourself whether your own cursor movements look human or bot-like.

The story so far

Until now, Cloudflare's bot defenses centered on "Risk": judging in one shot how likely a given request or action was to be harmful. According to the post, Risk is evaluated "in the moment," symbolized by one-time checks like CAPTCHAs, essentially a point-in-time assessment.

There was also a bot list called the "Bots Directory," but it only listed "good" bots, so bad bots, or bots whose identity wasn't yet clear, had to be handled separately.

But on today's internet, the line between a human operating something and an AI agent operating it has gotten blurry enough that it can flip mid-session (the post calls this the "Agentic Internet"), and a single point-in-time judgment risks missing that kind of shift.

What changes

From here on, things shift to "Trust," a continuous evaluation based on reputation built up over time.

• BotBase replaces the Bots Directory, which only listed good bots, with a ledger that records both good and bad bots together
• It checks what bot operators declared about themselves against how they actually behave
• Precursor injects JavaScript from the CDN side, so it keeps watching behavior for the whole session with no extra setup needed on the site's side

The post also spells out the conditions for being "Verified": meet both of these, and you're treated as a "good bot."

• You honestly declare who you are
• You don't abuse the trust you've earned

Break either one, and you get flipped back to unverified.

For site owners, the big change is being able to judge trust from behavior across an entire session, instead of deciding to block or allow based on a single request. And since anyone can try Precursor Trace, even people who aren't engineers can intuitively experience how their own cursor movements get evaluated.

Dive Deep

Let's go a bit deeper on the technical side.

Precursor's biggest feature is that it evaluates behavior "continuously throughout the session." The post puts it this way:

Precursor evaluates user behavior continuously throughout the session, so no more free hall passes.

As a 24-hour snapshot, the post shares these numbers:

• Evaluation events: 206 million
• Zones using Precursor: 73,438

Watching that volume of sessions continuously is what makes it possible to catch things a one-shot judgment would miss, like behavior shifting from human-like to agent-like and back again mid-session. The post also explains that this kind of continuous evaluation drives up the cost for bot developers trying to keep mimicking human behavior, making it economically less worthwhile.

The "Precursor Trace" tool you can actually try visualizes fine-grained traits people rarely notice about their own mouse movements, like:

• How your cursor accelerates
• Your habits around correcting your movement
• The rhythm and texture of your movement

I found it kind of fun to imagine trying it myself and wondering, does my own cursor movement look mechanical?

The post also previews two upcoming additions:

• Adaptive Intelligence Engine: an engine that automatically adapts and self-adjusts to behavior changes happening on the scale of hours or even minutes, without waiting for a formal version release. It's planned to roll out to all Bot Management customers
• Advanced Mitigations (rolling out by year-end): more ways to respond to suspicious traffic once it's detected. The post lists three approaches
  • Unpredictability: randomly returning block, challenge, or allow
  • AI Labyrinth: three modes, Maze (an endless chain of linked pages), Summary (useless AI-generated content), and Poison (deliberately fake data)
  • Queuing: managing the throughput of legitimate automated traffic

The post doesn't spell out detailed configuration screens or API names yet, but it does link out to the existing BotBase, Verified bots, and AI Labyrinth docs, so it's worth checking those if you're considering adopting this.

Wrap-up

To sum up this announcement:

• Cloudflare shifted how it judges bots and agents, from a one-shot "Risk" assessment to continuous, session-wide "Trust" evaluation
• "BotBase," a ledger tracking both good and bad bots together, replaces the old Bots Directory
• "Precursor," a detection engine that watches behavior continuously from the CDN side, processed over 206 million events across 73,438 zones in the last 24 hours
• "Precursor Trace" lets anyone try right now to see how their own cursor movements get evaluated
• By year-end, new responses like the Adaptive Intelligence Engine and Advanced Mitigations (Unpredictability, AI Labyrinth, Queuing) are also coming

This is worth a look whether you're an engineer working on security or bot management, or you're just curious where the internet is headed as AI agents keep multiplying. If you're intrigued, go check out Precursor Trace and see how your own cursor movements get scored!