php2shell: AI-Driven WAF Blocked the Attack Before the Patch — The Age of Handwritten Rules Is Fading

On October 21, 2026, Calif.io, Wiz, OpenAI, and Quarkslab jointly disclosed two critical remote code execution vulnerabilities in PHP (CVE-2026-109618 and CVE-2026-109619), nicknamed “php2shell.” The impact covers every site running PHP 5.4 or later — and PHP underpins nearly 70% of all websites on the internet. An attacker needs no login: a single carefully crafted upload request that puts PHP’s memory under pressure can crash the service, leak sensitive data, or execute arbitrary attacker code.

What truly shook the industry wasn’t the vulnerability itself — a decades-old defect — but the speed with which it was weaponized. Wiz’s researchers put it plainly: frontier AI models have dramatically lowered the barrier from a longstanding bug to a working exploit. Mandiant’s M-Trends 2026 report adds a chilling data point: attackers now begin exploiting vulnerabilities an average of seven days before a patch is available. Worse, php2shell details leaked on October 20, before the coordinated disclosure, and attack traffic was already observed before the official fix shipped.

In other words: by the time a dev team reads the advisory, writes the WAF rule, and ships it, the exposure window is long gone.

On the defense side, AI was already moving first.

Cloudflare deployed php2shell WAF protections for all customers — including free-plan users — at 3:02 AM UTC on October 21, roughly 9 hours before public disclosure, automatically and with zero manual action. Notably, the defense doesn’t just pattern-match the request. It introduced a new “full transaction detection” capability: within a single HTTP transaction, it analyzes both the request and the origin server’s response to judge whether the exploit actually succeeded, then blocks the origin response before it reaches the attacker. This drastically reduces false positives against legitimate users. Cloudflare also disclosed that it observed attackers continuously mutating payloads to evade the rules, with its analysts updating the rules in real time.

F5 took a different path: no rules at all, just let the AI decide. F5 engineers took the public proof-of-concept exploit, built multiple variants — different encodings, payload placements, HTTP methods, and common evasion techniques — and fired them at an application protected by F5 WAF in blocking mode with AI-powered WAF enabled. No php2shell-specific rule was added; no configuration was changed. Result: the neural network, trained on global real-world attack traffic, rated the malicious requests 95/100 — critical risk — and blocked them on the spot, while normal requests to the same application passed through. F5 even revised its guidance: turn on detection at all three confidence levels (high, medium, low) and block anything the AI rates high or medium risk — let the model handle false-positive filtering.

Both approaches point to the same thing: the rhythm of security defense is shifting from “humans patching rules after the fact” to “AI judging in real time.” The response window is compressing from days to minutes, even seconds. Security teams once kept new rules in “log-only, don’t block” mode out of fear of false positives, waiting for human confirmation before acting — but in the face of machine-speed attacks, that buffer is exactly the time the attacker gets to use.

For SMBs, this is the hardest part. Big vendors have global security teams watching the changelog 24/7 and can push protections within hours. But most small and mid-sized businesses have no dedicated security engineer, no overnight on-call, and no way to react in the same hour a vulnerability leaks. Vulnerabilities don’t care about your size: the attacks that target PHP won’t skip your site just because it’s small.

Making “AI real-time judgment + virtual patching” a normalized, always-on 24/7 auto-defense is the real survival path for a website of typical scale — defense speed determined by AI, not by a human shift schedule.