Cloudflare has introduced Adaptive Intelligence, a new detection engine built into Bot Management that continuously learns from real traffic and adjusts its defenses as bot techniques evolve. The company wants to attack one of the advantages cybercriminals have today: automating attempts cheaply, constantly testing variations, and adapting their tools until they find a way around the protections.
Cloudflare Adaptive Intelligence in 30 seconds
- Adaptive Intelligence plugs into Cloudflare Bot Management and continuously retrains its models on real traffic.
- Cloudflare analyzes more than a trillion requests a day to detect automation patterns.
- The system will generate temporary, campaign-specific rules that later disappear.
- It will also combine network signals with behavior observed across the entire session via Precursor.
- Some of these capabilities are still rolling out gradually and aren’t fully available on day one.
The approach starts from a shift in how defenses work. Instead of relying solely on relatively stable rules that an attacker can study for days or weeks, Cloudflare wants certain parts of its protection to change fast enough that the work needed to dodge them stops being worth it.
The company frames the problem in economic terms. Building automated campaigns keeps getting cheaper thanks to residential proxies, networks of compromised devices, tools freely available online, and AI systems that make it easier to configure and modify attacks.
The defender is in a different position: it has to block the attacker without blocking real users.
Adaptive Intelligence is designed to narrow that asymmetry.
A model that stops waiting for the next update
Bot-detection systems typically combine several signals.
Cloudflare already uses machine learning, browser fingerprints, behavior, heuristics, and mechanisms to recognize legitimate bots, such as search engine crawlers.
Bot Management currently assigns every request a Bot Score between 1 and 99. Low scores are usually tied to automated traffic, and businesses can use that value to build rules in their web application firewall or in Workers.
Adaptive Intelligence adds a new engine behind that process.
The first capability available is continuous retraining of the machine-learning model.
Instead of preparing a version, validating it, releasing it, and waiting for the next update, Cloudflare says the system can absorb patterns from new tools and evasion techniques as they show up in traffic.
The company analyzes more than a trillion requests a day looking for signs of automated abuse. That volume gives it a considerable distributed dataset for observing how bots change across different sites and networks.
The goal is to shorten the gap between a technique appearing and getting folded into the defenses.
That doesn’t mean any new behavior automatically gets flagged as malicious. A system like this specifically has to avoid letting faster adaptation also drive up false positives.
Cloudflare says new configurations are validated before being applied to live traffic, though actual behavior will depend on each environment and the signals available.
Rules built to disappear
The most eye-catching part of Adaptive Intelligence will be the generation of disposable, temporary rules.
The logic is different from permanently keeping a rule designed to block one specific technique.
If an attacker figures out which characteristic is triggering the block, they can modify their bot to get past that check. They then hold onto that knowledge and reuse it.
Cloudflare wants to stop that learning from being worth much.
Adaptive Intelligence is built to generate rules targeting a specific campaign or pattern, keep them active for a limited period, and then retire them.
How long they last can vary on purpose.
The intent is to keep the attacker from having a stable target they can analyze repeatedly.
| More traditional defense | Adaptive Intelligence’s approach |
|---|---|
| Relatively stable rules | Temporary rules |
| Version-based updates | Continuous retraining |
| Mostly per-request signals | Request and session signals |
| The attacker can study the response | Tries to reduce the information the attacker gets |
| Protection changes periodically | Part of the defense can change continuously |
There’s an important caveat: this capability isn’t complete at launch.
Cloudflare explains in its technical documentation that continuous retraining is the first component being activated now, while automatic generation of disposable rules and other features will roll out later.
So the announced vision is broader than what’s actually available on day one.
Hiding from the attacker what’s been detected
Another part of the approach is reducing the information available to whoever is trying to get around the protection.
When a defense always responds the same way to a given signal, an attacker can run tests to figure out which aspect of their behavior triggers the block.
Adaptive Intelligence tries to avoid that direct feedback loop.
Cloudflare explains that the engine can detect a signal without necessarily reacting to it in a visible way. That way, the bot can keep using a trait the platform has already identified without knowing it’s been spotted.
The system also isn’t meant to rely on a single binary condition.
It uses a statistical judgment built from multiple signals, which makes it harder to isolate one specific rule and tweak only the parameter needed to beat it.
This approach doesn’t make evasion impossible.
Cloudflare itself starts from the opposite premise: a sufficiently persistent attacker can find a way through a given defense.
The strategy is to make that solution stop working quickly and force the attacker to redo the work.
Catching attacks that look human on every single request
Another current problem is that some bots try to behave like normal users.
They can use real browsers, residential IP addresses, deliberate pauses, and movements designed to mimic human interaction.
Each request looked at in isolation can seem legitimate.
The pattern shows up once you look at the full session.
Cloudflare recently introduced Precursor, a browser-based verification system that analyzes visitor behavior over time.
Precursor runs code on the client, collects signals throughout the session, and continuously updates the score assigned to the visitor.
Adaptive Intelligence will combine that information with telemetry pulled from across Cloudflare’s network.
Signals the company mentions include TLS JA4 fingerprints, request structure, challenge results, network reputation, session behavior, and data from Turnstile and Precursor.
That makes it possible to look for a category of threats that’s especially hard to spot: so-called low and slow attacks.
A credential-stuffing attempt, for example, can deliberately slow down to stay under typical request thresholds.
Something similar happens with certain scraping systems.
If each address only makes a handful of requests, a defense based purely on volume might not flag it as suspicious.
Analyzing activity across different time windows makes it possible to spot patterns that only show up when you look at aggregate behavior.
AI is also making the attacker’s job cheaper
Cloudflare ties Adaptive Intelligence to the growing use of artificial intelligence on the offensive side.
Attackers don’t need to have fully handed their operations over to autonomous agents for AI to have an effect.
Generating scripts, tweaking configurations, analyzing errors, mimicking specific interactions, or adapting a tool now takes less expertise than it did a few years ago.
There are also commercial services that offer proxy infrastructure or automation to users who don’t need to technically control every component themselves.
The result is a drop in the cost of running lots of tests.
For an online store, a financial institution, or a platform with user accounts, that ease can translate into constant waves of scraping, automated account creation, promo abuse, or login attempts using stolen credentials.
The difficulty is that blocking more aggressively can hurt real customers.
A residential IP address, for instance, is no longer enough on its own to assume there’s a real person behind it.
Adaptive Intelligence tries to use Cloudflare’s scale to spot traits that repeat across campaigns even when each individual attack tries to look different.
Automated defense also has to keep false positives in check
A protection system that automatically modifies its own models and rules introduces another risk: blocking legitimate traffic by mistake.
For an e-commerce business, mistaking a customer for a bot can have an immediate cost.
Cloudflare says new updates can be tested against real traffic before being fully rolled out, in order to check their accuracy without causing disruption.
The company also uses its customers’ traffic as a signal to improve the system. When a classification turns out to be wrong, that information can later feed back into training.
This capability has an advantage inherent to running a global network: a pattern spotted on certain sites can provide useful information for others.
But that doesn’t make detection foolproof, either.
Normal behavior varies a lot between applications. An API, an online store, a financial service, and a content site can all have completely different legitimate automation running against them.
Cloudflare specifically maintains categories for verified bots and lets customers decide which actions to apply based on the Bot Score and the type of traffic.
The next front will be telling good bots, bad bots, and AI agents apart
The rise of AI agents makes this classification even more complicated.
Not all automated traffic is malicious.
Search engines need to crawl pages. Monitoring systems make automated requests. Business services rely on APIs. And AI agents can access a site to look up information or carry out a task explicitly requested by a person.
Cloudflare is already adapting its tools to distinguish between different kinds of AI bot behavior, including crawlers built for search, training, and agents acting on behalf of users.
That’s why Adaptive Intelligence can’t just settle the question of whether a request came from a machine.
The question becomes what kind of automation is acting, how it behaves, and what it’s trying to do.
That’s an important distinction.
A heavy scraping bot, an assistant checking a page on someone’s behalf, and a search engine crawler can all use similar technology, but they represent very different scenarios for the site owner.
Combining behavior, reputation, global telemetry, and session signals is meant to provide more information for making that call.
Cloudflare wants to make adaptation the attacker’s cost
The most interesting idea behind Adaptive Intelligence isn’t really about using machine learning to detect bots. That’s already been part of Bot Management for a while.
What’s new is deliberately designing the defense to change before the attacker can get a return on the work spent getting around it.
It’s a race against time.
An attacker finds a technique that works, tests it, scales it, and gets a financial return. If the defense stays stable long enough, that effort pays off.
If the technique stops working shortly after being developed, they have to start over.
Cloudflare wants to keep shrinking that window.
It’s still an open question how much of this cycle can be automated without driving up false positives or introducing hard-to-audit behavior. Several of the features that make up the full Adaptive Intelligence vision are also still awaiting rollout.
What the announcement does reflect is a broader shift in cybersecurity.
Defenses can no longer assume offensive techniques will evolve at the pace of monthly or quarterly updates. Attackers can test variations continuously and use automation to speed up that process.
Cloudflare’s answer follows the same principle: a defense that observes, learns, changes, and re-validates on an ongoing basis.
The fight against bots is shifting from finding one perfect rule to making sure no evasion technique stays useful for too long.
Frequently Asked Questions
What is Cloudflare Adaptive Intelligence?
It’s a new detection engine built into Cloudflare Bot Management that continuously learns from traffic and adapts its models to new automation and evasion techniques.
Does Adaptive Intelligence already generate temporary rules automatically?
Cloudflare has initially launched the continuous-retraining component. Automatic generation of disposable rules and other features described by the company will roll out gradually.
How does Adaptive Intelligence relate to Precursor?
Precursor analyzes visitor behavior throughout the entire browser session. Adaptive Intelligence can combine those signals with Cloudflare’s global network telemetry to improve detection of hard-to-spot automation.
Who can use Adaptive Intelligence?
Cloudflare says Enterprise Bot Management customers can access the new system through the machine-learning auto-update option. Availability of specific features may evolve as new components roll out.