AI-Found PHP Flaw Forced Global Patch Scramble, and Cloudflare Counted Millions of Exploit Attempts

A flaw in PHP, the programming language behind as much as 70% of the web, was patched and publicly disclosed on Wednesday, Oct. 21. By then, word was already out.

Cyber agencies in Europe, the U.S., the U.K., Australia and Canada posted urgent warnings. The NSA told any site running PHP to patch immediately. Government and private websites in the U.S. and U.K. went dark temporarily while teams installed the update. No major hacking damage was immediately apparent.

Three days later, the story is less about the damage than about how close the system came to failing.

How the flaw was found

On July 30, researchers at the AI security firm Calif, working with OpenAI's Patch the Planet initiative, pointed a large language model at PHP. They did not expect to find anything critical. On Aug. 1, they found a way to attack it. More followed in the days after.

In early August, Sagi Tzadik, a researcher at Wiz, asked an AI agent to test whether a WordPress flaw could be turned against Drupal. Eight hours later the agent said it had failed. It then added that it had found a flaw in PHP. Tzadik assumed the model was hallucinating. "Models tend to hallucinate," he told TIME. The exploit worked, and the weakness had gone undetected for more than a decade.

Wiz and Calif had been working separately. They joined forces after both sent reports to PHP. About 15 people from competing firms spoke almost daily in shared channels, with no contracts governing the arrangement, to trace the vulnerabilities and prepare a fix.

CNN traced the underlying weakness class back further. Researcher Stefan Esser warned about a critical PHP flaw in July 2004. A partial fix came in 2020 after another researcher found a related problem. A similar vulnerability surfaced in 2024 and was left unfixed. CNN's account is that exploiting it by hand was too tedious to be practical, and AI worked out how in two days.

The leak

The patch was scheduled for later in the week. According to TIME, word of the vulnerabilities began circulating online two days before the release date, and PHP's maintainers moved the patch up by a day. Hours later, with the patch still unavailable, Calif mistakenly posted a video of its exploit at work. Copycat attempts multiplied.

Cloudflare counted thousands of attempts before the public announcement, tens of thousands within 45 minutes of it, hundreds of thousands within two hours, and eventually millions. "If it had leaked a month earlier," before a patch was ready, "this would have been a very different situation," said Tomer Schwartz, a Wiz chief technology officer.

Separately, Chainguard CEO Dan Lorenc released information about the flaw ahead of the embargo, which was meant to give defenders time to build a fix. He apologized in a LinkedIn post. In an interview with POLITICO on Thursday, Lorenc said he shared it early because the flaw was already being exploited, and that he had seen enough evidence to invalidate the embargo.

Others dispute that. Lorenc also said he does not believe his actions caused problems for other defenders. He conceded that the evidence threshold for breaking an embargo in these coordinated systems is a "grey area."

The sources do not say how much of the exploit traffic traces to the leaked video, to Lorenc's disclosure, or to the circulating rumors.

The warnings from tech's top ranks

On Thursday evening, OpenAI's Sam Altman said, "I no longer think we're going to be able to patch all bugs."

Jason Clinton, Anthropic's deputy chief information security officer, predicted a "vulnerability tsunami." "There won't be enough time to patch critical infrastructure," he wrote on X. Salesforce CEO Marc Benioff called it a "code red."

Cloudflare CEO Matthew Prince wrote that "the next two years are going to be incredibly, incredibly, incredibly fraught with new tools that have infinite patience to go look for vulnerabilities across the Internet."

Vyas Sekar, a Carnegie Mellon computer engineering professor, said there are more PHP-like vulnerabilities out there, some known and many unknown. "This may be the first of many. This is a wakeup call."

The month has picked up the label "Black October." Eylam Milner, co-founder and CTO of the security startup Echo, called the PHP flaws "about as serious as it gets for the web." Wiz described it as one of the most severe vulnerabilities in software history.

The case that the system worked

The people who found the bugs make a different argument. Tzadik says the episode shows defenders winning. "We are fixing vulnerabilities at an unprecedented rate, and this is doing security overall a very good thing," he told TIME.

The design behind that outcome is the one TIME describes. AI companies give defenders early access to new models before the same capabilities reach malicious users. This time, Wiz and Calif got there first.

That is a voluntary, private-sector arrangement, and it produced a patch before the exploit went public. It also depended on a handful of people keeping a secret for weeks, and the secret did not hold.

What comes next

The threat is no longer theoretical. The Guardian reports that attempts to exploit the PHP flaw have already begun, and the NSA warned that attackers could use these bugs to access servers without authorization.

The unresolved question is the one Lorenc's "grey area" comment points to. Nobody has set a clear standard for when a disclosure embargo can be broken. With Altman, Clinton and Prince all predicting more flaws of this size, the next one will test that gap again.