OpenAI, Google Wiz AI Tools Uncover Decades-Old PHP Security Flaws

Full analysis

Two AI security efforts, OpenAI's "Patch the Planet" and Google's Wiz, found two critical flaws in PHP that had sat in the code since 2004. PHP runs something like 70% of the web. The NSA told anyone who can't patch to pull their site offline until they can. The same AI that found the holes, Wiz warns, now lets attackers adapt the exploit across versions and configurations. So the question for anyone running or buying AI tools: is machine-assisted vulnerability discovery a real inflection, and which side of the attack line does it help more?

This is easy to undo for the reader as news, but the underlying shift is not. You can patch PHP this week. You cannot un-invent the capability that found the bug, and the next finder may not call the PHP Foundation first.

The Skeptic

Start with the model names, because they don't check out. "GPT-5.6 Sol" and "Gemini 4 Argon" are not shipping products anyone can point to. That's either sloppy reporting or a press release that got copied as fact. Treat the whole technical claim as unverified until the CVE scores land and an outside researcher reproduces the method. "Flaw exists since 2004" is a headline, not a severity rating. Half the ancient CVEs were obscure because exploiting them needed conditions that almost never line up. And "70% of websites" counts millions of dead WordPress blogs with nothing worth stealing. The gap between "AI found a 20-year-old bug" and "AI found something attackers will actually use at scale" is wide, and nobody has closed it here.

The Safety Lens

The dual-use part is the real content, and for once it's concrete, not a thought experiment. The Wiz quote spells it out: LLMs can port an exploit across versions, configurations, and memory layouts. That porting work used to be the thing that kept obscure bugs obscure. It took a skilled human days. Now it takes a prompt. Responsible disclosure worked this time because two companies with reputations to protect chose to make it work. There is no rule forcing the next finder to coordinate with the PHP Foundation instead of selling to a broker. The NSA's "go offline" language is an admission that patching speed can't keep up with discovery speed. That's a structural problem, and patching PHP doesn't touch it.

The Researcher

Twenty years of open-source eyeballs on PHP missed what these tools surfaced in days. If that holds up, the signal isn't "PHP was buggy," it's that AI-assisted code review now finds flaw classes human auditors skip. The obvious follow-on: PHP is not special. The same C-era foundations run the internet, OpenSSL, glibc, imagemagick, zlib, all decades old, all picked over by humans who missed things. If the method generalizes, we're at the front of a long queue of "ancient flaw found" stories. The responsible-disclosure coordination here is the template. What's missing is a structured triage program, because ad hoc initiatives named "Patch the Planet" won't scale to the volume this is about to produce.

The Compute Pragmatist

The economics are the part nobody's pricing honestly. Running continuous LLM-assisted analysis deep enough to find a 2004 bug is not one clever prompt. It's sustained, context-heavy inference over huge codebases, and at today's token prices that bill is real. OpenAI's "Patch the Planet" branding tells you they're absorbing the cost, so the true number stays hidden. That matters because the defender's advantage right now rests partly on OpenAI and Google eating a cost an attacker couldn't. As inference gets cheaper through distillation (training a small cheap model to copy a big one) and better chips, always-on vulnerability scanning drops into reach for people with no margin to protect and no disclosure ethics. The balance tips when the cost of scanning falls below the payoff of a working exploit.

The Builder

If you run PHP in production, this is your week, and the NSA's "pull it offline" is not boilerplate. The trap is that PHP lives under WordPress, Magento, and Laravel apps nobody has touched in four years, where "patch now" meets "we're scared to redeploy this." So the move is a web application firewall rule or virtual patch before your patch cycle finishes, because the window between a patch shipping and a weaponized exploit appearing has compressed. The adaptation capability Wiz flagged is exactly what closes that window. Assume your instance is exploitable, not merely misconfigured, until you've proven otherwise.

Where they split

The Skeptic and the Researcher are the real fight. The Skeptic says prove it: unverifiable model names, a misleading reach stat, no published severity, so this reads as much like an AI marketing event as a security one. The Researcher says the method is the story, and PHP is just the first of many. Both can't be fully right. If the CVE scores come in high and someone reproduces the technique, the Researcher wins and we're at the start of a wave. If the scores are mild and the method doesn't replicate on the next codebase, the Skeptic called it.

The second split: the Safety Lens and the Compute Pragmatist agree the balance tips toward attackers, but disagree on when. Safety says the capability already exists and only goodwill held it back here. Compute says the goodwill is propped up by cost, and the cost is falling fast. Either way, the thing protecting defenders today is not a technical moat. It's two companies choosing to behave and a bill attackers can't yet pay.

What it hinges on

Whether this is an inflection or a press cycle comes down to one thing you can check: do the CVE severity scores publish high, and does an independent team reproduce the discovery on a different old codebase? That's verifiable. Everything else, the reach stat, the drama of "since 2004," the model version strings, is narrative until then.

Prediction: Before 2027-04-30, at least one more critical vulnerability in a widely-used, decades-old open-source project (such as OpenSSL, glibc, imagemagick, or zlib) will be publicly disclosed with AI-assisted discovery named as the method.

Confidence: Medium. The method generalizes and both labs are racing to repeat the PR win.

Why: OpenAI's "Patch the Planet" and Google's Wiz both branded these PHP finds as flagship security wins, which means both have a running incentive to point the same tools at the next big legacy codebase and publicize what they find. The technique that surfaced a 2004 PHP flaw, deep LLM-assisted static analysis, is not PHP-specific. The same C-era foundations run the internet and have had the same human-only review for decades. The opposite outcome, that nobody announces another AI-found flaw in six months, would require both labs to stop running a program they just launched and publicized, which cuts against their own incentive to own the "AI defends the internet" narrative.

Revisit by 2027-04-30: We're right if a critical CVE in a major decades-old open-source project is disclosed with AI-assisted discovery credited as the method. We're wrong if no such AI-attributed critical disclosure in a comparable project appears by that date.

The quieter read: the first time one of these finds gets weaponized before the patch propagates, the "AI defends the internet" story flips to "AI broke the internet" overnight, and the same labs taking victory laps now will be the ones explaining themselves.