Intel / blog

When scanners miss the attack: how Cloudflare Client-Side Security protects storefronts

Source: blog.cloudflare.com
Date: Published Sep 16, 2026

Context

Cloudflare reports finding eight malicious JavaScript payloads across four storefront campaigns using its Page Shield machine-learning system. Some scripts waited for particular visitor conditions before acting, so an apparently normal page load could conceal unwanted behavior. The company's investigation describes commission diversion, click interception and conditional loading of further code.

Its detector examines code structure, including relationships between symbols and calls, to recognize suspicious patterns despite renaming or some obfuscation. A checks flagged scripts before customer alerts. The mechanism gives detection a basis beyond whether a script's address or exact contents already carry a malicious label.

Key takeaways · AI-distilled

Cloudflare's Page Shield ML system caught all 8 malicious JavaScript payloads it studied across 4 e-commerce fraud operations, while VirusTotal was missing 7 of the 8 and URLScan flagged none of them as malicious.
The detection engine is a graph neural network that reasons over JavaScript as a syntax-tree graph of code symbols rather than flat text, which lets it spot suspicious patterns despite minification, renaming, or obfuscation without needing a known signature.
Under 0.3% of analyzed traffic gets flagged by the GNN for a second opinion from a lightweight LLM on Workers AI; the hardest cases go to an ensemble of about six frontier 'teacher' models whose votes are weighted by their score on the Artificial Analysis Intelligence Index.
One affiliate-hijacking script used a MutationObserver to detect dynamically loaded product buttons, then hijacked clicks to route sales credit to the attacker via a typosquatted domain, adtargett[.]com instead of adtarget[.]com.
A mobile-only cloaking payload checked a 325-entry IP denylist plus device, geography, and campaign-tag gates before firing, then deleted scripts for nine analytics tools and swapped in the attacker's own ad and analytics IDs.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source blog.cloudflare.com