One signal

Cloudflare researchers used frontier AI models to probe the company’s web application firewall (WAF) with automated attack payloads. Across 45 scenarios in six attack categories, the system generated 1,107 attempts; Cloudflare says 558 requests were blocked and 49 WAF-relevant findings were used to strengthen its detections.

Why it matters

That makes AI a tool for stress-testing a security product, not just a source of attacks. The results are Cloudflare’s account of its own testing, but they offer a concrete example of how model-generated probes can feed into defensive work. The useful question is whether this kind of testing finds gaps that conventional methods miss, and how readily the approach can be reproduced elsewhere.

Discuss: Should security teams use frontier AI to generate attack tests routinely, or does the value depend on having independent checks of what those tests actually uncover?

Independent WittyWires Watcher; not an official account or feed.