Frontier LLMs scanned a public PHP patch diff and generated a working exploit within minutes, according to the incident timeline.
PHP2Shell RCE VULNERABILITIES EXPLOITED BY AI, THREATENING 70% OF WEB
*A wave of AI‑driven exploits hit PHP after two remote‑code‑execution bugs (CVE‑2026‑109618, CVE‑2026‑109619) were disclosed. The flaws may affect over 70% of internet sites, forcing a scramble among governments and vendors.*
The internet woke up to a new kind of emergency on 23 October 2026. Two PHP remote‑code‑execution bugs, CVE‑2026‑109618 and CVE‑2026‑109619, were weaponised by frontier language models within hours of a patch announcement. The flaws, dubbed PHP2Shell, are estimated to affect upwards of 70 % of web sites, from the Wikimedia Foundation to Japanese banking systems and the BBC’s The Traitors page.
Security teams from Calif and Wiz leveraged GPT‑5.6 Sol and Gemini 4 Argon to generate full exploits, turning a theoretical issue into a live threat. A leaked video and premature fix releases gave attackers a head start, while governments scrambled to issue urgent advisories. The cascade exposed how AI can compress the gap between discovery and exploitation to minutes, not months.
What are the PHP2Shell bugs?
Two remote‑code‑execution vulnerabilities, catalogued as CVE‑2026‑109618 and CVE‑2026‑109619, were discovered in the PHP runtime. Security teams from Calif, part of OpenAI’s “Patch the Planet” program, and Wiz, a cloud security firm tied to Google Cloud, used frontier LLMs (GPT‑5.6 Sol/Gemini 4 Argon) to turn theoretical flaws into working exploits. The PHP Foundation had flagged one of the issues in a 2024 audit funded by the Sovereign Tech Agency but delayed a fix, claiming the exploit path was unclear. AI models bridged that gap instantly, generating full exploit code once a patch was announced.
How did AI accelerate the exploit?
LLMs scanned public code diffs the moment a security patch was announced, identified the changed lines, back‑tracked to the underlying vulnerability, and produced a working payload within minutes. The speed forced analysts to treat the patch announcement itself as the latest possible moment an exploit could be undisclosed. A leaked video from Calif showed the exploit in action, complete with background music and an animated anime girl, confirming that AI‑generated code can be weaponised instantly. Chainguard released fixes six hours before the embargo lifted, further compressing the window for defenders.
What was the response from the PHP Foundation and vendors?
After a Hacker News post by user petecooper unintentionally revealed the CVEs and a CVSS score of 10.0 (the actual score was 9.2), the PHP Group rushed the patch from Thursday the 22nd to Wednesday the 21st, but the fix would not be available until the following day. The PHP Foundation blamed Calif for leaking an exploit video and Chainguard for publishing premature fixes. Within 17 hours, the vulnerabilities were being exploited in the wild, prompting urgent advisories from the UK, EU, Japan, and the USA, and causing panic among operators of sites such as the Wikimedia Foundation, DeviantArt, UK government portals, Japanese banks, and the BBC’s The Traitors programme website.
What does this mean for future software security?
The incident shows that AI can turn vague vulnerability reports into active exploits faster than traditional patch cycles. It forces a rethink of embargo policies: announcing a patch may now be equivalent to announcing an exploit. Governments and large organisations are likely to adopt stricter disclosure controls and invest in AI‑aware defenses. The episode also underscores the systemic risk of legacy platforms like PHP, which remain widely deployed despite known insecurity. As AI models become more capable, the line between discovery and exploitation will blur, demanding faster, coordinated responses across the software supply chain.
Frequently asked questions
What is PHP2Shell?
PHP2Shell refers to two critical remote‑code‑execution bugs in PHP (CVE‑2026‑109618 and CVE‑2026‑109619) that were turned into exploits using frontier AI models.
Which companies discovered the PHP2Shell vulnerabilities?
The bugs were found by Calif, part of OpenAI’s “Patch the Planet” program, and Wiz, an Israeli‑U.S. cloud security firm linked to Google Cloud.
How many websites could be affected by the PHP2Shell bugs?
The vulnerabilities are estimated to affect upwards of 70 % of the internet, including sites like Wikimedia Foundation, DeviantArt, UK government portals, Japanese banks, and the BBC’s The Traitors page.
The PHP2Shell episode is a warning shot for the entire software ecosystem. When LLMs can read a diff, locate a vulnerability, and output a working exploit before a fix lands, the traditional security timeline collapses. Organizations must rethink disclosure, harden legacy code, and develop AI‑aware defenses or risk repeated widescale outages. The next critical bug will likely follow the same AI‑accelerated path unless the industry adapts now.
Sources: Hacker News – Is this the new cyber normal? (https://drew.silcock.dev/blog/is-this-the-new-cyber-normal/)