Two severe remote code execution vulnerabilities in PHP—CVE-2026-109618 and CVE-2026-109619—were discovered by security teams at Calif and Wiz using frontier AI models. According to the source, these bugs are estimated to affect upwards of 70% of the internet, impacting thousands of major sites including Wikimedia Foundation, DeviantArt, UK government services, Japanese banking systems, and the BBC’s The Traitors website.
One of the vulnerabilities had been identified during a 2024 security audit by the PHP Foundation, funded by the Sovereign Tech Agency, but was initially overlooked because no one could develop a practical exploit from the theoretical issue. The AI tools proved effective at converting vague vulnerability descriptions into functional exploits.
The coordinated disclosure process broke down when a Hacker News user named petecooper posted about the CVEs on October 20, 2026, apparently unaware of an embargo. The PHP Group accelerated the patch release from October 22 to October 21, but the patch still wouldn’t be available until the following day. According to the source, the PHP Foundation blamed Calif for leaking a video demonstration of the exploit and Chainguard for publishing fixes six hours before the embargo was set to lift, allowing attackers to move faster than system administrators.
The source notes that AI’s speed in converting security patches into working exploits has fundamentally changed the timeline for vulnerability response. It states that “the time from ‘hey I heard there’s a bad RCE in this project’ to ‘there are sophisticated, automated, and concerted attacks’ has gone from weeks/months to hours/minutes.” The vulnerabilities were being actively exploited in the wild before legitimate security patches reached most users.
Governments worldwide, including the UK, EU, Japan, and USA, issued urgent advisories recommending immediate action, amplifying public panic. The source suggests the anxiety stemmed partly from PHP’s reputation for security issues and partly from the genuine severity of these particular vulnerabilities. However, the source also notes that many LLM-generated exploits may be non-functional “slop,” requiring genuine expertise to produce working attacks.
Key facts
- Two critical remote code execution vulnerabilities in PHP discovered using AI models, estimated to affect 70% of the internet
- One vulnerability had been identified in a 2024 audit but dismissed as unexploitable until AI tools built working exploits
- Responsible disclosure broke down when the vulnerability was publicly disclosed before embargo, accelerating exploitation timeline from weeks to hours
- The PHP Foundation, Calif, and Chainguard faced accusations over the timing and handling of vulnerability disclosure and patch release
- Multiple government agencies issued urgent security warnings, contributing to widespread panic