AI helped turn old PHP bugs into critical flaws, and attackers moved within hours

On 21 October 2026 the PHP project released emergency fixes for two critical vulnerabilities, known together as PHP2Shell. Both can allow unauthenticated remote code execution on servers that run PHP. According to Wired, citing analysis from Cloudflare, millions of exploitation attempts arrived within hours of the flaws becoming public.

The flaws themselves are not new. Parts of the underlying problem were documented years ago and left unfixed because nobody could show they were exploitable. Researchers using AI models showed that they were. Wired's framing is that AI turned a historic vulnerability into a current crisis. The PHP Foundation says it has not seen any successful exploitation so far, according to Wired, and many of the exploits now circulating appear not to work.

The bugs

The PHP security release covers two CVEs, both rated 9.2 under CVSS 4.0. CVE-2026-109618 is remote code execution through out-of-memory corruption. CVE-2026-109619 is remote code execution through stale global state. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP's advisory says there are no complete mitigations and that "updating is the only safe choice". It names mod_php under Apache as the highest-risk setup, "with a demonstrated exploit where an unauthenticated request can trigger the attack before any application code runs."

The common cause, as Calif describes it, is that PHP can leave dangling pointers or inconsistent internal state when a request is cut short by a memory-limit error or another fatal error. Code that runs later can then pick up that stale state. Calif says the earliest affected code path dates to PHP 5.4.0. It also says the exploits work under default configurations and reports pre-authentication code execution against WordPress, Drupal, Joomla, MediaWiki, Magento and Nextcloud, among others. Wiz says every PHP version released in the last decade is affected by at least one of the issues.

A long history, then a proof

Calif's timeline goes back two decades. In 2004 the researcher Stefan Esser published a remote code execution technique based on PHP's memory limit. A related bug, #78876, was reported to PHP in 2020. In 2024 an audit by OSTIF and Quarkslab found a related out-of-memory issue. Calif says it was not fixed then because exploitation had not been proven.

Calif found several exploitable paths on 1 August 2026. It reported the first bug to PHP on 11 August, shared a working proof of concept on 14 August, and reported the second bug on 21 August. Wired reports that the work was part of OpenAI's Patch the Planet program. OpenAI and Wiz later found PHP2Shell flaws on their own. Wiz dates its first finding to 19 August.

"The vulnerability had been known to the PHP team for years, but because it could not be leveraged by an attacker, at least, not a human one, it was not addressed," Wiz researcher Sagi Tzadik told Wired.

What the AI did and did not do

The accounts from the researchers are more modest than the headlines. Calif says it used no special harness and no restricted models. Its researchers chose the target, began from Esser's 2004 technique and searched for surviving variants. They used a language model to review allocation and shutdown code. The model pointed out a counter that was incremented before an allocation that could fail. The team then turned that pattern into a skill file so agents could search for similar bugs. Calif says language models do not reliably start from obscure, decades-old primitives unless someone points them there. Ryan Austin, the Calif researcher who first found the flaws, wrote: "No model is simply capable of unearthing 'all bugs' with simple prompts. At least for now."

Wiz says it used its own Wiz Atlas tool together with Gemini 4 Argon and GPT-5.6 Sol. It credits Gemini 4 Argon with an "invaluable contribution" to finding the vulnerability, building the exploit, detecting attacks and validating the patch. Wiz also says AI helped adapt the exploits to different PHP versions, configurations and memory layouts by analysing failed attempts and trying again. That kind of iteration used to be slow, specialist work.

The researchers Wired spoke to expect more cases like this, but not a flood. Jake Williams of Hunter Strategy, a former NSA hacker, said there will be more "long-life bugs" but "only so much of this low-hanging fruit", because few other tools are exposed to the internet the way PHP is. Security engineer Niels Provos said he is "not really worried": people have been searching with AI for more than a year, and "it required a highly skilled team to find this one."

What it means for work

For security researchers, AI now covers much of the tedious middle of the job: reading large codebases, generalising a bug pattern and adapting an exploit to many targets. Choosing where to look still takes human expertise. For maintainers, "not proven exploitable" is no longer a safe reason to leave a known bug open, since a model may supply the missing proof, and the backlog of old reports becomes a list of risks. For administrators and IT teams, the time between disclosure and attack is now measured in hours. Wired notes that Microsoft's updates this month alone covered a record 1,453 vulnerabilities, and that Bank of America hired 1,000 consultants to fix flaws found by Anthropic's Mythos. It is still open whether AI will uncover many more old bugs that can be exploited at scale, or whether PHP's reach makes this case unusual, as Williams and Provos suggest.