PHP crisis exposes the internet’s fragile open-source foundations

TechCrunch’s Kirsten Korosec estimated that PHP (Hypertext Preprocessor) sits behind content management systems used by about 70% of websites, putting tens of millions of sites in the potential path of a vulnerability emergency. The estimate turns a language disclosure into an internet-scale maintenance problem.

The TechCrunch Equity episode opened with an announcement from earlier in the week: website owners had to patch PHP immediately or, if they could not update in time, take their sites offline. Korosec described PHP as a core ingredient in WordPress and Drupal, where many owners may not know their sites contain the decades-old language at all.

Sean O’Kane called the episode “a real reckoning moment.” He said that “when you get down to the bone level of the internet,” everything seems “held together by glue and dreams,” a warning about both fragile open-source dependencies and the tools now available to find flaws.

How did a decades-old flaw become an emergency?

Anthony said the vulnerability was not new. It had existed in PHP for decades, and nobody knew about it before researchers disclosed the problem, he said.

That silence matters because PHP has been used to build much of the web for decades. A flaw in a foundational layer can therefore remain dormant until a disclosure forces many unrelated operators to respond.

The scale pushed operators toward visible shutdowns. O’Kane said governments took websites down, while popular websites also went offline briefly so engineers could fix the vulnerability. The practical result was emergency downtime wherever patching could not happen immediately.

Did AI find the PHP vulnerabilities?

Anthony initially described AI companies, including Wiz, now owned by Google, as finding the flaws with large language models (LLMs), the advanced software O’Kane said researchers could now task at scale. He then corrected that account: Human researchers participating in OpenAI and Wiz’s Hack the Planet initiative found the vulnerabilities, although they used AI to exploit the bug.

The correction matters because it narrows the claim from autonomous AI discovery to human-led research with AI-assisted exploitation. The TechCrunch Equity episode did not identify the individual flaws, affected PHP versions, or a firm patch deadline, so it documented urgency rather than a complete technical chronology.

O’Kane said Google and OpenAI had people search for vulnerabilities, including PHP, while attackers moved to exploit the disclosures. That creates a security dilemma: silence can protect users temporarily, but only public disclosure gives maintainers a route to ship a fix.

The practical rule is straightforward: researchers found the bug, AI helped exercise it, and disclosure exposed defenders and attackers to the same instructions. Any broader claim that AI alone discovered the PHP flaws would go beyond the episode’s corrected account.

Why did disclosure trigger an exploit race?

O’Kane said many people disliked the disclosure rollout and saw chest-beating in OpenAI’s post about the flaws. He also said a couple of people got ahead of the embargo, the publication restriction, which set off a race among hackers.

Cloudflare CEO Matthew Prince posted that he was seeing the problem escalate to millions of attempted hacks of PHP2Shell, according to O’Kane. Korosec called Prince’s post both telling and scary.

The normal cyber-disclosure path starts with a researcher reporting a problem to the vendor, O’Kane said. PHP was too broad for that quiet process because it helps run many different computers and parts of the internet, he said, so broad notice created an unavoidable race. The episode reported millions of attempts, not a matching count of successful compromises.

O’Kane said the easier identification of webmasters might have solved part of the notification problem, but only if AI were advanced enough to task a billion agents with emailing every site controller. He called this the beginning of repeated back-and-forth and a new cybersecurity frontier. The unresolved question is who can notify and patch more sites before the next disclosure window closes.

Who can keep up with the exploit wave?

Amazon also published a blog post about the emergency, according to Korosec. The company operated a honeypot network, or a set of decoy systems, and detected exploitation attempts before the public announcement, she said.

Amazon had the tools, capital, team and know-how to monitor those attacks, Korosec said. Many other websites, including potentially government sites, lacked that level of attention, she said. The disparity makes the threat worse because weaker operators may remain exposed after better-defended companies have moved.

Her second concern was link rot, meaning links that break when sites disappear. Companies, governments and even small neighborhood sites may choose shutdown over patching, leaving historical pages inaccessible and weakening the public record.

Korosec also called the episode just the beginning, asking what other decades-old flawed software might be exploited next. She warned of the “next level of prepper doom […] because you could see how banks could get exploited, and [with] data and privacy concerns, it just feels a lot bigger than we can fix right now.”

The immediate gap is defensive capacity, not merely awareness of the bug. Amazon could monitor decoys before disclosure, while many site owners lacked equivalent resources.

What could AI agents change next?

O’Kane opened the broader argument by saying the internet was breaking in several ways that week. He described a recurring pattern over the previous 10 years: one person maintaining a small code repository as a hobby catches signs of malicious activity and prevents wider infrastructure from crumbling. His multiplier is a billion, because people can now task LLMs and AI agents—software tasked with pursuing a goal on its own—to search and act.

The scale is a forecast, not evidence that AI found the PHP flaws on its own. The episode’s researchers were human, even though they used AI during exploitation.

Korosec asked how extreme people’s behavior would become and whether new startups would emerge. She also pointed to open-source security tools offered by companies such as Nvidia as possible protective layers.

The TechCrunch Equity episode reported no successful bank compromise, no measured benefit from AI-assisted defense, and no complete inventory of aging software. It also did not show that agents caused the initial PHP disclosure. Its evidence supports a narrower claim: shared software remains difficult to disclose, patch and monitor at the same pace as attackers can test it.

The episode established a repeatable sequence: researchers disclose shared flaws, exploit attempts race against patches, and less-resourced operators absorb disproportionate risk. The next test is whether coordinated disclosure and patching can stay ahead of attackers before aging infrastructure forces another round of emergency shutdowns.