13 Nov 2026

AI Didn't Just Patch PHP2Shell — Wiz Says It's What Made the Exploit Possible in the First Place

Wiz Research says frontier AI models were essential to weaponizing PHP2Shell, a bug class humans found too complex. What that means for risk.

When PHP shipped emergency patches for the PHP2Shell vulnerabilities (CVE-2026-109618 and CVE-2026-109619) on October 21, 2026, we covered what website owners needed to do about it at the time. What that advisory didn’t yet have was Wiz Research’s own account of how the vulnerabilities were actually found and turned into a working exploit — published the same day, and worth a separate look specifically because its real subject isn’t PHP. It’s what AI-assisted vulnerability research is starting to do to the gap between “theoretically exists” and “practically exploitable.”

What Wiz Research actually disclosed

  • Same two CVEs, now with a named research story behind them. CVE-2026-109618 and CVE-2026-109619 — memory-safety bugs in PHP’s core engine, not any specific application — carry a CVSS score up to 9.2 and affect nearly every PHP version released in roughly the last decade, back to PHP 5.4.
  • A genuinely huge affected footprint. Because PHP sits underneath WordPress, Drupal, Joomla, and MediaWiki (the software that runs Wikipedia) among countless custom applications, Wiz’s own estimate puts the realistic exposure at roughly 70% of websites — not a metaphor, a rough measure of how much of the public web runs on some version of the affected runtime.
  • The root cause: how PHP mishandles fatal errors and out-of-memory conditions during request processing, leaving corrupted internal state that a later, separate request can then reuse — opening the door to unauthenticated remote code execution, or a Heartbleed-style disclosure of raw process memory that can include passwords, session tokens, and API keys.
  • The part that’s actually new here: Wiz’s researchers say frontier AI models were essential to building a working exploit chain for a bug class that had been theoretically present in PHP for roughly a decade, but had been too complex for human researchers to practically turn into a functioning attack.
  • A fast, coordinated response. PHP shipped a patch the same day as disclosure, October 21, 2026, alongside mitigation coordination involving Automattic, Akamai, AWS, Cloudflare, Fastly, Google, Imperva, Microsoft, and Vercel, among others.

The real story isn’t the PHP bug — it’s the research timeline

Strip away the specific CVE numbers, and the structurally interesting claim is this: a vulnerability class that sat in one of the most widely deployed pieces of software on the internet for roughly ten years, unexploited not because nobody looked but because actually weaponizing it was too complex for manual human research, went from “theoretically present” to “working exploit” once frontier AI models were brought into the research process. Wiz’s own framing compares this moment to Heartbleed and Log4Shell as a symbolic turning point for an “AI era” of vulnerability research — that specific comparison is the researchers’ own characterization, worth repeating as their view rather than as settled industry consensus, since Wiz is also a vendor with a commercial interest in positioning its research as significant.

What doesn’t depend on how dramatically anyone frames it: the practical cost of turning a complex, long-standing theoretical vulnerability into a working exploit appears to be dropping, because AI-assisted research can now do work that used to require rare, specialized, time-intensive manual expertise. That’s a real, checkable shift in the underlying economics of vulnerability research, independent of whether this specific case deserves a Heartbleed-sized headline.

Why this changes the risk conversation, not just the patch conversation

Every security program has, at some point, implicitly or explicitly accepted a risk because “it would take serious, rare, specialized effort to actually exploit this” — a legitimate, standard part of risk triage when expert human research time is genuinely scarce and expensive. The PHP2Shell research story is a concrete data point suggesting that assumption needs re-examination: if AI-assisted research can compress a decade of impractical complexity into a working exploit chain in a research team’s working timeframe, “too complex to realistically exploit” is a weaker argument than it was a year ago, for any class of complex, long-standing vulnerability — not just this one.

This connects directly to what we’ve covered in AI red teaming and LLM security testing: the same AI capability that compresses a defender’s testing timeline compresses an attacker’s exploit-development timeline too, and there’s no reason to assume only one side of that equation benefits. It’s also the dual-use framing we apply consistently on this blog, covered in AI-generated malware — the same capability is genuinely valuable for defensive research (which is exactly how Wiz used it here, for responsible disclosure) and genuinely risky in the hands of an attacker, and treating it as only one or the other misses the actual shape of the problem.

Security recommendations

  • Stop treating “too complex to practically exploit” as a durable risk-acceptance justification. Revisit any security decision in your own environment that currently rests on that assumption, specifically for long-standing, complex vulnerability classes rather than simple, well-understood ones.
  • If your organization does any internal security research or code auditing, evaluate AI-assisted tooling for your own defensive use — the same capability that compressed PHP2Shell’s research timeline is available to defenders doing legitimate, authorized testing, not just to whoever gets there first.
  • Patch on the assumption that disclosed-but-complex vulnerabilities get weaponized faster than they used to. The operating assumption that a complex bug gives you a comfortable patching window because exploitation is hard is less safe than it was.
  • Re-evaluate any software with a long support history for classes of bugs that were previously dismissed as impractical to exploit — a decade-old, widely-deployed codebase is exactly the profile AI-assisted research is now well-suited to revisit.
  • If you haven’t already, confirm your own PHP version is patched — CVE-2026-109618 and CVE-2026-109619 themselves are covered step-by-step in our original PHP2Shell advisory; this article is about the research story behind them, not a replacement for that remediation guidance.

AI-era vulnerability risk checklist

  • Identify any current risk-acceptance decisions that rely on “too complex to exploit in practice” as their justification
  • Evaluate AI-assisted security research/code auditing tools for your own authorized, defensive use
  • Re-review long-standing, widely-deployed software in your stack for vulnerability classes previously dismissed as impractical
  • Confirm patching timelines assume a compressed window between disclosure and real-world exploitation, not the old assumption that complexity buys time
  • Confirm your own PHP version is patched against CVE-2026-109618 and CVE-2026-109619 specifically

Key takeaways

The patch for PHP2Shell closed a specific, serious vulnerability pair — that part was already covered. What Wiz Research’s own disclosure adds is a research story worth taking seriously on its own terms: a bug class that sat unexploited for a decade because it was genuinely too complex for manual human research became a working exploit once AI entered the process. Whether or not it deserves the Heartbleed comparison Wiz gives it, the underlying shift — AI compressing the distance between “theoretically vulnerable” and “practically exploited” — is a real one, and it applies well beyond this one PHP bug.

Shahid Security’s Vulnerability Assessment engagements specifically re-examine long-standing “it’s too complex to realistically exploit” assumptions in a client’s environment, not just scan for already-known, easily-weaponized issues.

Have legacy or long-unpatched systems you’ve previously judged too complex to be a realistic target? Book a consultation.


Sources: Wiz Research — PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet, our own PHP2Shell advisory and remediation guide.

FAQ

Is this a new vulnerability, separate from the PHP2Shell CVEs already patched?

No — this is the same CVE-2026-109618 and CVE-2026-109619, which PHP patched in emergency releases on October 21, 2026. What's new is Wiz Research's own disclosure write-up, published the same day, which focuses specifically on how the flaws were found and turned into a working exploit — not a new flaw. If you still need the patching and remediation steps, see our original PHP2Shell advisory.

Did Wiz actually build a working exploit, or just describe a theoretical attack?

Wiz Research says they built a working exploit chain as part of their own security research, used to confirm and responsibly disclose the vulnerability to the PHP project — not deployed against live third-party systems. The point of their write-up is specifically that building this working exploit required frontier AI assistance; by their own account, the underlying bug class had been present in PHP for roughly a decade without a practical human-built exploit.

Is it fair to call this a turning point comparable to Heartbleed or Log4Shell, like Wiz does?

That comparison is Wiz's own framing in their research write-up, not an independently verified industry consensus — and Wiz, like any security vendor, has some incentive to frame its own findings as dramatically significant. Worth noting plainly rather than repeating as settled fact. What's independently checkable is the underlying mechanism: a real, scored, patched vulnerability pair affecting a huge share of the web, which by the researchers' own account required AI assistance to practically exploit after a decade of being theoretically present. That part is worth taking seriously regardless of how the comparison is framed.

Shahid Iqbal

Written by Shahid Iqbal

Founder & Cybersecurity Consultant, Shahid Security