Two critical vulnerabilities in PHP, the programming language behind a large share of the web, are being broadly exploited after they were disclosed on Wednesday, October 21. The flaws, nicknamed PHP2Shell, were found with the help of frontier AI models from OpenAI and Google and had gone unnoticed for more than a decade, according to the Financial Times report carried by Ars Technica. The US National Security Agency told organizations running PHP to update immediately or take affected systems offline. It is a concrete, high-stakes example of AI finding bugs at internet scale, and of attackers using the same capabilities.

What was found

The two flaws sit in PHP itself, not in a single website or plugin. The NSA described them as often giving attackers “a direct path to the server’s file system or database.” The UK’s National Cyber Security Centre said they could let an attacker take control of a website’s server without a password or any other login credentials. PHP’s maintainers released patched versions on Wednesday.

Researchers at two US security groups, Calif and Wiz, are credited with the discovery. Wiz, which is owned by Alphabet, estimated that PHP powers about 70 percent of websites worldwide. A large part of that footprint comes from WordPress, the most widely used website platform, which is written mainly in PHP.

How AI models were involved

OpenAI said its GPT-5.6 Sol model tracked down the flaws as part of an initiative it calls Patch the Planet. The company called the issue one of the “most impactful security issues in recent memory” and said that, to its knowledge, frontier AI found the vulnerabilities before attackers did. Wiz reported that it found the same flaws independently using its own Wiz Atlas platform together with Google’s Gemini 4 Argon and GPT-5.6 Sol. According to Wiz, Gemini 4 Argon helped with finding the bug, developing exploits, building detection for attacks and validating the patch.

The models did more than spot a bug. Wiz argues that language models lower the cost of turning one working exploit into many. An exploit that worked against one PHP installation used to need significant manual effort to adapt to other versions, configurations and memory layouts; models can now do much of that adaptation automatically. In Wiz’s words, flaws “that once seemed impractical to exploit become viable targets.”

Exploitation began before disclosure

Amazon said on Wednesday that it had seen an attempt to exploit the weakness four days before the public disclosure, and that the flaws were “now being broadly exploited in the wild.” Ars Technica’s headline frames this as a four-day head start for attackers. We have not seen independent figures on how many sites are compromised, and the scale of exploitation so far rests on statements from Amazon and government agencies rather than a public count.

Governments moved quickly. The NSA and UK authorities issued warnings, the White House said federal agencies were assessing the effect on government systems and critical infrastructure, and some European governments were reportedly taking sites offline while they patched.

A disclosure process built for AI-speed findings

The fix was coordinated over two months in five phases. Participants included the PHP developers, Wiz, Cloudflare, and US agencies including CISA and the White House, according to The Decoder’s summary of OpenAI’s announcement. OpenAI is proposing the process as a repeatable framework for future AI-driven security discoveries. It also expects findings of similar size to “increase sharply in the coming months” and warns that the same capabilities can help attackers write offensive tools quickly.

That warning matters because the traditional pattern, where a researcher finds a bug, notifies the vendor and waits for a patch, assumes that the number of serious bugs found per month is limited by human effort. If models can find and weaponize flaws faster, defenders need shorter patch cycles and better coordination. A tool published on GitHub lets server operators check whether their PHP version is affected.

What to watch next

  • Whether a public count of compromised sites emerges, and how fast hosting providers and WordPress site owners apply patches.
  • Further details from PHP’s maintainers, CISA and the NSA on indicators of compromise.
  • Whether OpenAI’s Patch the Planet process is adopted by other vendors and labs.
  • How regulators respond; earlier this year, powerful cyber-capable models from Anthropic and OpenAI prompted warnings from politicians and financial regulators and a temporary limit on rollout by the US government, according to the FT report.
  • Follow-up disclosures from AI-assisted researchers in other widely used open-source projects.

Key takeaways

  • Two critical PHP flaws, nicknamed PHP2Shell, were disclosed and patched on October 21, 2026.
  • OpenAI’s GPT-5.6 Sol and Google’s Gemini 4 Argon helped researchers find them; the flaws had existed for more than a decade.
  • Amazon says exploitation began four days before disclosure and is now broad; the NSA says patch immediately or take systems offline.
  • Wiz estimates PHP runs about 70 percent of websites, largely through platforms such as WordPress.
  • The same AI that finds bugs also helps attackers adapt exploits, so patch speed matters more.

Frequently asked questions

Do I need to act if I run a WordPress site?

Check with your hosting provider or administrator which PHP version the server uses and install the patched release published on October 21. The NSA’s guidance is to update at once or take the system offline until it is mitigated.

Who found the vulnerabilities?

Researchers at Calif and Wiz are credited, working with OpenAI and, in Wiz’s case, Google’s Gemini 4 Argon. OpenAI says its model located the flaws before attackers did.

Does this mean AI will make software less secure?

It cuts both ways. The same models that helped defenders find and patch these flaws can help attackers adapt exploits. The outcome will depend on how quickly vendors patch and how well disclosure is coordinated.