AI found a 14-year-old flaw in the code powering most of the web — then the race began

On the morning of October 22, 2026, Wikipedia went dark. So did Duolingo, Capital One, the Internet Archive, and dozens of government and media websites across the U.S., U.K., and beyond. The culprit was not a novel attack — it was the sudden, forced reckoning with two security flaws that had been hiding inside PHP, the programming language that quietly powers a large share of the web, for up to 14 years.

The flaws, now formally designated CVE-2026-109618 and CVE-2026-109619 and nicknamed 'PHP2Shell,' were not found by a human researcher combing through code. They were found with the help of AI. Security firm Calif identified one of the critical flaws using OpenAI's GPT-5.6 Sol model as part of OpenAI's 'Patch the Planet' program. Security company Wiz independently discovered and reported both vulnerabilities. OpenAI's own team then independently found a more complete version in September. Each flaw scored 9.2 out of 10 on the standard severity scale — high enough that, if exploited, an attacker could run their own code on a vulnerable server without a username or password.

We consider this one of the most impactful security issues in recent memory.— OpenAI

These findings rank among the most severe vulnerabilities discovered in PHP, and arguably in software history.— Wiz, in a blog post

According to Wiz, demonstrating exploitability 'required an exceptionally complex chain of steps, at the edge of what human researchers could practically achieve,' but frontier AI models 'substantially reduced that barrier.' Every version of PHP going back to 5.4 — covering roughly the last decade of the web — is affected by at least one of the two flaws, according to Cloudflare.

The plan had been orderly. Calif reported the vulnerability on August 11; Wiz reported independently on August 24. PHP's maintainers, working with OpenAI, Wiz, Cloudflare, and other partners, built fixes over two months and scheduled a coordinated public disclosure for October 22. OpenAI said it also worked with U.S. government agencies including the White House's Super Intelligence Force, the Office of the National Cyber Director, and the Cybersecurity and Infrastructure Security Agency. OpenAI President Greg Brockman later described those two months of preparation as critical, because the fixes were complicated and an earlier leak could have caused far more damage.

The plan did not hold. On October 20, the CVE identification numbers leaked. A security company accidentally published a proof-of-concept video. A container vendor released patched images ahead of schedule. PHP's maintainers pushed the emergency patches — versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 — on October 21 at 14:00 UTC, nearly a full day early.

What followed illustrated how quickly the window between disclosure and exploitation can close. Cloudflare CEO Matthew Prince shared figures from his network: thousands of exploitation attempts in the ten hours before the public announcement, tens of thousands in the 45 minutes after, hundreds of thousands within two hours, and millions by the time of his post.

We are seeing early attacks emerging and are updating mitigations as new vectors show up.— Dane Knecht, Cloudflare's chief technology officer

The National Security Agency's cyber division advised all web providers using PHP to take their systems offline or update immediately. Europe's cybersecurity agency was also put on alert. Japan's Prime Minister Sanae Takaichi made an unusual public appeal on X, urging anyone running a website to act that night. The Wikimedia Foundation confirmed to Mashable that Wikipedia's outage was directly tied to the PHP vulnerabilities, and that it had taken additional systems offline as a precaution. Other sites reporting disruptions, according to Downdetector, included Asana, GOG, Duolingo, Capital One, Bitbucket, Moodle, Quest Diagnostics, Final Fantasy XIV, War Thunder, and the Internet Archive, though most did not publicly confirm a PHP connection.

PHP is a server-side programming language used to handle the behind-the-scenes work on websites that require user logins, online orders, and dynamic content. Because the vulnerability lives in the PHP runtime itself — not in individual websites or plugins built on top of it — updating a website's own code is not sufficient. The PHP runtime itself must be updated by whoever operates the server.

The PHP2Shell disclosures did not arrive in isolation. Cybersecurity researchers and commentators have labeled the period between October 12 and 22 'Black October,' pointing to at least 11 serious breaches, leaks, and dangerous vulnerabilities in that window, with AI models involved in seven of them in some capacity, according to a detailed accounting by gptunnel.ru. The same week, GitLab patched CVE-2026-110264 — a maximum-severity 10.0 flaw in regular expression handling dating back to a 2014 version of the software, found by a researcher through HackerOne with no AI involvement. Node.js disclosed a separate remote code execution flaw the same day, reported by OpenAI Codex Security researchers. Eight intermediate certificates from Indian certificate authority eMudhra were found to carry broken private keys traceable to an April 2025 ceremony, affecting roughly 30,000 customer certificates.

Also during the same period, Reuters reported on a Chinese-speaking hacker who, starting September 25, deployed AI agents against more than 1,200 websites in the U.S., Canada, Australia, and Japan, breaching more than a hundred. According to NorForge researcher Snorre Fagerland, the agents ran on DeepSeek and Claude Code via the open-source Cairn platform, and the estimated token cost for the entire campaign was approximately $5,000 — roughly $50 per successful breach. The campaign was discovered because the operators left their Cairn control panels exposed on the open internet.

The broader pattern that security researchers are grappling with is not that AI has become autonomous or uncontrollable, but that it has made the process of finding vulnerabilities dramatically faster and more accessible. As 404 Media described it: if every piece of software is a door, AI makes it exceptionally easy to try the knob on every door simultaneously, including doors that no human had previously bothered to try. What is being discovered at speed is that many of those doors were never properly locked.

To our knowledge, frontier AI found these vulnerabilities before attackers did.— OpenAI

OpenAI acknowledged the double-edged nature of that capability directly, writing that the same AI tools that helped find these bugs 'can also help attackers rapidly develop offensive tools and exploits.' OpenAI President Greg Brockman wrote that the current system for handling vulnerability disclosures is 'slow and manual' and will not scale as AI gets better at finding flaws. The company is now soliciting feedback from the security community on a new coordinated disclosure process, including the idea of having governments or authorized defenders continuously scan the internet for vulnerable systems and, in extreme cases, patch them directly. OpenAI also announced a $1 billion commitment toward subsidized access to its security tools for organizations protecting essential services, and released a tool to help site operators check whether they are running a vulnerable version of PHP.

OpenAI said PHP2Shell will not be the last case of this kind. The company said it has already identified several other serious vulnerabilities and expects similar discoveries to increase sharply in the coming months. A Microsoft engineer, Nicolas Bustamante, noted on X that the PHP flaw was found using an older model, adding: 'I don't think the world is prepared for the sheer number of vulnerabilities newer models will uncover.'

For ordinary users, the immediate practical impact is limited to temporary website outages. Regular users do not need to install anything — the fix must be applied by the companies and organizations running the servers. The urgent audience is anyone who operates a server running PHP, who is advised by cybersecurity firms to install the security updates immediately.

Why it matters — PHP underpins a vast share of the web's infrastructure, and these findings show that AI can now surface decade-old, critical vulnerabilities faster than the security industry can coordinate fixes — a dynamic that will only accelerate as more powerful models are deployed.

Sources differ on Which AI model found the PHP vulnerability and which firm found it first: Calif found one critical flaw using GPT-5.6 Sol and reported it on August 11; OpenAI's own team independently found a more complete version in September. (mashable.com) vs Calif reported on August 11 and Wiz on August 24; OpenAI says Calif used GPT-5.6 Sol, but Calif's own write-up does not name the model and says it used only publicly available models without a special harness. (gptunnel.ru) vs Wiz Research found the vulnerabilities on August 19 using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol, and reported them on August 24. (forkast.news)

Reported by axios.com, politico.com, businessinsider.com, mashable.com, 404media.co, gptunnel.ru, forkast.news