Cloudflare's new bot defense rewrites its own rules mid-attack

Renting a bot army is now nearly free. Updating the defenses against one still takes weeks. Cloudflare CTO Dane Knecht framed the mismatch in a single line: "building taller walls fails when the cost of scaling an attack is effectively zero."

That cost gap is what the company's new Adaptive Intelligence engine is built to attack. The detection layer inside Cloudflare Bot Management writes and discards its own blocking rules while an attack is in progress. The rules are narrow, short-lived signals, rotated faster than the attacker can map them.

Conventional bot-management products ship on scheduled updates, with new signatures landing weeks or months after a campaign starts to register. Attacker tooling turns over on a different clock: a few hours, sometimes less. Free scraping software and rented residential device fleets keep failed attempts effectively free at volume, so a defender who only patches on a release schedule is paying a slow, regular price to fight a fast, near-zero one.

Adaptive Intelligence inverts the cadence. Instead of waiting for a vendor release to ship a new rule, the engine watches the traffic moving across Cloudflare's network, more than a trillion web visits a day by the company's count, and writes candidate rules on the fly. New detections test themselves against live traffic in the background, checking accuracy and screening for false positives before deploy, with no downtime for the protected site. The rules that go live are narrow and short-lived by design; the system rotates them so an attacker probing the target is mapping a defense that has already changed.

The engine also catches what scheduled signatures miss. It analyzes behavior across several timeframes at once, which lets it surface slow campaigns hidden under ordinary rate thresholds: drawn-out credential stuffing runs and scraping jobs that look like a normal user on a per-second basis but are not, in aggregate, behaving like one. Operators that run below a per-second threshold are the population a per-second signature update cannot reach, and the engine's multi-timeframe analysis is built to find them.

Adaptive Intelligence pulls from Precursor, the browser-level telemetry system Cloudflare made generally available in July. Precursor logs movement, scrolling, and typing across an entire session and feeds that into the bot engine as a richer signal of who is at the keyboard. Cloudflare said in July that bots already accounted for roughly 57% of web requests when Precursor reached general availability, a baseline figure for how much of the traffic Adaptive Intelligence is being asked to read. It is also the first major bot-management product to ship on top of a vendor that has already published a bot-traffic baseline at that scale, which gives security teams a fixed number to compare future measurements against rather than the usual "more bots than last year" hand-waving.

The cost-asymmetry argument holds only if rotation outpaces bypass engineering. Adaptive Intelligence ships short-lived rules rather than permanent signatures, so an attacker who reverse-engineers a block today has a smaller shelf of rules to reuse tomorrow. Cloudflare has not published third-party benchmark data for the engine, so the cost-asymmetry argument and the rotation cadence are company-stated rather than independently measured. The first independent tests of the rotation cadence and the false-positive rate, when they land, will be the numbers worth tracking, and a third-party measure of how long a given rule stays effective in the wild is the falsifier that decides whether "moving target" is a marketing term or a design property.