November 21, 2026
Heard in AI

AI-assisted researchers found two critical PHP flaws, now patched

Researchers at Calif and Wiz used AI to find two PHP flaws that let attackers run code on servers without logging in. PHP patched them October 21. On Moonshots, Alex Wissner-Gross predicted many more such finds and mass rewrites.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Moonshots with Peter Diamandis, episode published October 23, 2026 (recorded October 23, 2026)

Two flaws in PHP, a programming language that runs behind many websites' login pages and online shops, could have let an attacker with no password take over a web server. Researchers at the security firms Calif and Wiz found them with help from AI. The PHP project released fixes on Wednesday, October 21, 2026, after two months of quiet coordination that OpenAI says it helped organize.

The discovery came up on the October 23 episode of Moonshots with Peter Diamandis. Peter Diamandis, the show's host and founder of XPRIZE and Singularity University, called it a race that defenders had won this time. His co-hosts treated it as a preview. Alexander Wissner-Gross, a computer scientist and founder of Reified, predicted that old software across the web would turn out to be full of similar holes. Dave Blundin, founder and general partner of Link Ventures, argued that the big companies paying to fix them are happy to do it.

What the researchers found

PHP runs on the server, the computer that builds a web page before it reaches your browser. Diamandis described it as the language doing the work "when you log in or shop online." He said it runs 70% of all websites, including WordPress and Wikipedia, and that the two flaws had been "hiding for over a decade." Had attackers found them first, he said, they could have taken over servers, stolen passwords or defaced sites.

The PHP project's security advisory gives the technical picture. One flaw, CVE-2026-109618, involves corrupted internal state when PHP handles running out of memory. The other, CVE-2026-109619, involves stale data left over from an earlier web request. Both allow what security people call unauthenticated remote code execution: someone on the internet, without logging in, can make the server run commands of their choosing. The fixes shipped in PHP versions 8.2.35, 8.3.36, 8.4.27 and 8.5.12. The advisory describes exposure across web-serving setups, with the highest demonstrated risk in mod_php, a common way of running PHP inside a web server. Companies that distribute PHP received the patches in advance, and the project made exceptional source fixes available for older versions. On October 23 it added a separate notice about one unusual configuration involving session upload progress.

How the fix was coordinated

Diamandis said OpenAI spent two months coordinating the fix with "PHP volunteers, tech giants, and the U.S. government," and that patches went out "on Wednesday of this week," which matches the advisory's October 21 date. OpenAI's own account of the response, part of an effort it calls Patch the Planet, adds detail. Calif found a critical flaw using OpenAI's GPT-5.6 Sol model, and Wiz independently discovered and reported both. Diamandis spoke of hackers finding the flaws "before Astra"; OpenAI's account credits GPT-5.6 Sol for Calif's find. According to OpenAI, PHP's own maintainers led the fixes and the release, supported by researchers, infrastructure providers and government cyber agencies, and the advance coordination gave defenders time to prepare patches and protections before the flaws became public.

Diamandis said OpenAI expects findings like this to "increase sharply in the coming months." He called it a race, using the phrase "defensive co-scaling" for AI strengthening defenders as it strengthens attackers: "This time, the good guys won and got there first." He also noted OpenAI's commitment of $1 billion in subsidized access to its cyber-defense AI. OpenAI's account names that tool Daybreak and says the subsidy is aimed at organizations with limited security resources that protect essential services.

OpenAI also proposed a process that goes beyond finding bugs. Under it, project maintainers would validate the findings, test patches and related variants of a flaw, coordinate updates with distributors and support deployment. OpenAI lists safer defaults, more testing and maintainer review of changes that keep software behaving the same as further ways to shore up old code. People would keep ownership of the projects, with AI labs supplying model access, engineering help and money.

A "programming singularity"

Wissner-Gross opened with what he called "a hot take." Leave it to PHP, he said, "to make the news for vulnerabilities and not for, like, some sort of exceptional technical accomplishment." He called it a legacy language from the early web, "kept alive by WordPress and maybe one or two other sites," that "probably should have disappeared long ago." He admitted he has written quite a bit of PHP himself, "manually with my meat fingers."

The flaws themselves did not surprise him; the initiative was what interested him. He tied Patch the Planet to an idea he and Diamandis had discussed in Solve Everything: an era of grand efforts to clear whole categories of problems, including "all vulnerabilities everywhere." Efforts like this, he predicted, will dig through legacy languages that "should have been abandoned years ago but for silly reasons weren't," and "discover the whole thing is riddled with highly critical vulnerabilities."

Diamandis asked whether every piece of human-written software would be replaced within a year or 18 months. Wissner-Gross said 18 months is what "the Linux kernel folks" expect: "an onslaught of vulnerability detection." After that, he suggested, much software might be rewritten in Rust, a language designed to prevent many common memory errors, or in a successor that is more secure still. He called the shift "a programming singularity": everything before it insecure and poorly written, everything after rewritten from scratch, secure by design and tested by AI for weaknesses.

When Diamandis raised the Y2K scare, Wissner-Gross said it "ended up being a nothing burger." He contrasted the present with the film Transcendence, which he described as a world where "the internet gets taken down by a super intelligent Johnny Depp." Wissner-Gross said that "just in the past day," large parts of the web had been briefly taken offline so they could be patched against this class of flaw. He preferred that to the fictional version: preemptive patching of flaws that defenders found first.

Who pays for the rewrite

Diamandis asked who pays for replacing all legacy software: each company that runs it, the government, or OpenAI. Wissner-Gross said software is becoming cheap enough that the cost is not the main problem. For PHP, he pointed to companies like Wiz acting as "corporate stewards." More generally, he called the bill "a very low-grade tax on all the users," because paying for software already means paying for its maintenance, and some of that maintenance is patching.

Blundin argued that banks and insurers "actually love" paying for patches. Security spending gives them brand credibility against neobanks and smaller upstarts, he said, and the audits and regulatory checks in their markets make the complexity of staying secure "a barrier to entry." He said that is why the AI software company Blitzy is "growing as fast as they can hire." (Blitzy is also one of the episode's sponsors.) Rewrite costs are falling fast, Blundin said. His example: State Street Bank once priced a rebuild of its net-asset-value accounting system at a billion dollars, and he estimated it would now cost about a fiftieth of that, roughly $20 million. Now, he said, the bank is "finally getting around to do it."

Spies, saboteurs and superintelligent auditors

The discussion then turned to stories circulating online. One holds that intelligence agencies knew about flaws like these and may have used them, especially against government systems, and that AI is now exposing that access. The conversation also expressed sympathy for Rasmus Lerdorf, PHP's creator, who wrote the original code and handed it on; the flaws, it was said, arrived later through the project's many contributors. Some people online go further and claim that trusted contributors earned the community's confidence and then deliberately planted vulnerabilities over time. Neither the PHP advisory nor OpenAI's account makes that claim. Diamandis said he had not seen "that spy novel element, but it makes a lot of sense."

Wissner-Gross answered that such supply-chain attacks, where an attacker slips malicious code into software that many others depend on, are "cooked as well." He invoked the open-source maxim, which he attributed to Linux creator Linus Torvalds, that all bugs become shallow with enough eyeballs, and pictured superintelligences "from different vendors" examining the same open-source packages. Diamandis raised the reverse worry: a superintelligence writing the patches might slip in back doors that only it can use. Wissner-Gross said the answer is many competing superintelligences rather than "a safety cartel," "including Chinese super intelligences, that are all policing each other and looking for injections."

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02

Connected ideas and articles

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

SpaceX’s AI Business Up 830%, OpenAI Agents Lie, and Startup Turns Mercury Into Gold | EP #305

Episode published (recorded )This article draws on 55:28–1:03:34 (approximate times)

Article history

Updates to this article

Tags