Cloudflare Adaptive Intelligence, explained

Cloudflare Adaptive Intelligence matters because it shortens the defender’s update loop without asking you to rebuild your bot rules. On August 31, 2026, Cloudflare made continuous model retraining available to Enterprise Bot Management customers, but the disposable rules and customer-feedback learning in the bigger vision are not live yet.

What Cloudflare Adaptive Intelligence actually is

Cloudflare already gives Bot Management customers a bot score, a number from 1 to 99 that estimates whether a request came from automation or a person. A score of 1 means Cloudflare is highly confident it is automated. A score of 99 means Cloudflare is highly confident it came from a human.

Adaptive Intelligence sits behind that score. It changes how the machine-learning model, the statistical system that turns traffic signals into a probability, gets updated.

Before this release, the model waited for fixed releases. Attackers could study a stable detector, change their fingerprints, spread requests across residential proxies, and keep each address below an obvious rate limit. The defender’s model stayed still until its next update.

Now the model retrains continuously on live traffic. New weights, which are the learned settings inside the model, are tested and deployed automatically. Your existing WAF rules and Workers logic still receive the same bot score, so the contract around the model does not change. Cloudflare’s launch page describes the operating loop as observe, train, deploy, and validate.

The validation step matters. A candidate model first runs in shadow mode beside the current model, scoring real traffic without affecting visitors. Cloudflare compares results such as challenge solve rates. A candidate that scores real people worse does not become the primary model.

The score itself still follows the same ranges:

Here is the catch. Cloudflare describes Adaptive Intelligence as a larger system with disposable rules and learning from customer corrections. At launch, only continuous retraining is running. Disposable rules that appear and retire at changing intervals, plus the feedback system that turns corrected mistakes into training signals, are still coming later.

Why this changes the defense loop

The useful change is time, not a new dashboard object. Cloudflare says it analyzes more than one trillion requests per day for automated abuse. Adaptive Intelligence can combine signals such as JA4 TLS fingerprints, request structure, challenge outcomes, session behavior, network reputation, Turnstile telemetry, and Precursor telemetry across several time windows.

That is aimed at a specific kind of attack. One request can look normal. Many requests, spread across addresses and sessions but moving through the same login or checkout flow, can reveal one coordinated operator. A detector that studies only each IP address or each request misses the shared behavior.

Continuous retraining lets Cloudflare fold a new bypass pattern into the scoring model without waiting for customers to pick a version and schedule a migration. Your WAF expression can remain unchanged while the model feeding it adapts.

The launch does not prove how much better detection becomes. Cloudflare publishes no detection-rate improvement, false-positive rate, precision, recall, or before-and-after benchmark. More traffic and faster retraining describe the mechanism. They do not give you a measured business result.

This release also has a narrow audience. The activation path is for Enterprise customers with the paid Bot Management add-on. Pro and Business customers can still see grouped bot traffic through their existing products, but they do not get the granular Enterprise Bot Management control described here.

Who can use it tomorrow

An ecommerce security lead protecting checkout

Inventory hoarding and checkout abuse often arrive through many proxies, with each source staying quiet enough to avoid a simple rate limit. The security lead can turn on model auto-updates, keep the existing checkout rule in place, and compare bot-score distributions with challenge outcomes before changing the action.

The payoff is a shorter model-update cycle without a WAF rewrite. The team still decides whether a low score should be logged, challenged, rate-limited, or blocked.

A SaaS authentication team handling credential stuffing

A SaaS team can apply the score to login and account-recovery routes, where slow distributed attempts are harder to spot request by request. The model can look across short bursts and longer patterns, while the application keeps using the same score field.

The practical win is continuity. Existing rules do not need a new variable or a migration. The team can monitor whether the changing model catches more suspicious sessions without increasing failed logins for real customers.

A ticketing or marketplace operator protecting scarce inventory

This operator needs different treatment for a public catalog, a reservation endpoint, and static assets. Enterprise Bot Management supports path-specific WAF rules, so the operator can challenge likely automation on the scarce-inventory path while excluding verified bots and static resources.

That keeps the response tied to business risk. A crawler viewing public pages is not handled like automation trying to reserve every available slot.

A regulated enterprise security team governing model changes

Automatic model updates remove version selection from the customer. That is convenient for a fast-moving attack and awkward for a change-control process that expects a named version and approval record.

The security team should treat the toggle as a governed control. Record when it was enabled, preserve Bot Analytics and request logs, document the WAF actions that consume the score, and ask the Cloudflare account team how model changes and rollbacks appear in the organization’s audit process.

How to enable it safely

Cloudflare’s Bot Management setup guide gives the current dashboard path. The work is short, but the observation step is where the judgment sits.

1. 1

   Confirm the entitlement

   Adaptive Intelligence is attached to Bot Management for Enterprise, a paid add-on that an account team adds to an Enterprise plan. If the dashboard still shows Add Bot Management, the zone does not have the required entitlement.

2. 2

   Turn on Bot Management

   Open the Cloudflare dashboard, choose the zone, go to Security Settings, filter by Bot traffic, open Bot management, and turn it on. Entitled zones have the product available, but the zone setting still needs to be enabled.

3. 3

   Enable model auto-updates

   In Bot Management, open Configurations. Select the edit control for Auto-updates to the Machine Learning Model and turn it on. There is no Adaptive Intelligence version to choose and no migration to schedule after that.

4. 4

   Inspect the rules that consume the score

   Cloudflare documents these starting expressions for definite and likely bots. They exclude verified bots and static resources.

   (cf.bot_management.score eq 1 and not cf.bot_management.verified_bot and not cf.bot_management.static_resource)

   (cf.bot_management.score ge 2 and cf.bot_management.score le 29 and not cf.bot_management.verified_bot and not cf.bot_management.static_resource)

   The expression identifies traffic. Your rule action determines what happens next. Start by reviewing where the rule applies and which customer journey it can interrupt.

5. 5

   Watch before tightening

   Use Bot Analytics and Logs to inspect score patterns and the outcomes of your challenges. Keep a special eye on login, checkout, account recovery, and reservation flows. A changing model can improve adaptation, but your conversion and access failures are the evidence that tells you whether the action is too aggressive.

The honest part

The biggest limitation is commercial. Cloudflare lists Bot Management for Enterprise as a paid add-on supplied through an account team, and the plan page does not publish a fixed price. You cannot calculate the return from a public per-request or per-zone number.

The second limit is control. Once auto-updates are on, Cloudflare deploys model weights rather than asking you to select each version. Cloudflare runs candidates in shadow mode and can pause or roll back its rollout, but you do not get a customer-facing version picker. Teams with strict model-change approvals need to settle that governance question before enabling the toggle.

The third limit is evidence. The mechanism makes sense for attacks that change faster than scheduled model releases, but there is no public benchmark showing how many more bots it catches or how false positives change. Treat the first weeks as an observation period in your own traffic, not as proof that every current rule should become a block.

Finally, the launch name is ahead of the shipped surface. Continuous retraining is real now. Disposable rules and feedback-driven learning are the next parts of the plan. They should not appear in a security review as controls you already have.

What to do now

Act this week if you already pay for Enterprise Bot Management and face changing bot patterns on login, checkout, account recovery, or scarce-inventory routes. Enable auto-updates, leave the score contract alone, and measure the effect before changing rule actions.

Wait if your organization requires a named model version, a customer-controlled rollout, or a numerical vendor benchmark before production use. Take those requirements to the Cloudflare account team first.

You are largely unaffected if you do not have the Enterprise add-on. Keep using the bot controls included with your current plan. This release does not create a self-serve Adaptive Intelligence switch for Free, Pro, or Business accounts.

For one practical explanation like this when a platform changes, join the newsletter.