AI-Found PHP Flaws Trigger Fixes and Service Disruptions

PHP maintainers released fixes on October 21, 2026, for two critical vulnerabilities that could let attackers run code on servers without authentication. OpenAI said the same day that security company Calif found one of the flaws using its AI model GPT-5.6 Sol. In Japan, business, childcare and government services paused operations as a precaution, and Prime Minister Sanae Takaichi urged organizations to prepare to update PHP.

The issue in PHP itself, called “PHP2Shell,” could leave internal state inconsistent while handling memory exhaustion or fatal errors, potentially enabling remote code execution (running code on a remote server) under certain conditions. WordPress said no update to WordPress itself was needed and asked users to check their hosting providers’ guidance and update PHP. Calif’s discovery was part of OpenAI’s “Patch the Planet” initiative to support open-source maintenance. Wiz also independently found and reported vulnerabilities, saying AI helped test and debug attacks that combined multiple issues and were adjusted to different memory layouts on each target.

PHP maintainers had identified signs related to the root cause in 2024, but response stalled because no specific attack path had been found, fixes posed compatibility challenges, and staffing was limited. Reports in summer 2026 clarified the impact, making a fix the top priority. PHP maintainers led the fixes and release, with research teams and security companies helping validate them and prepare defenses.

According to the PHP Foundation, vulnerability identifiers and other details leaked online at around midnight Japan time on October 21, moving the planned release from October 22 to October 21. At around 5 a.m. that day, Calif mistakenly published a proof-of-concept video demonstrating a working attack; copies continued to circulate on social media after it was deleted. The foundation said the video accelerated development of proof-of-concept code. At 4:41 p.m., software distribution company Chainguard published images containing the fix more than six hours before the coordinated release was scheduled. The foundation said attackers could then analyze the fix.

In Japan, Chatwork carried out precautionary maintenance from the evening of October 21 into the early hours of October 22, while childcare and education services provider Codmon halted all services. Codmon advised facilities to use alternatives such as attendance and departure logs, and said in its October 21 notice that it had found no unauthorized access or information leaks linked to the vulnerability. Some Japan Meteorological Agency weather-data pages became inaccessible on the morning of October 22, though Kikikuru, typhoon information and rain-cloud forecast pages remained available, FNN reported. On the evening of October 21, Takaichi asked operators to check whether their public websites used PHP and to prepare to apply the fix scheduled for release at 11 p.m. that day.

Amazon Threat Intelligence said on October 21 that records from MadPot, a system that attracts and observes attacks, showed an exploitation attempt dated October 17. It said the route by which the attack information was obtained and whether the attempt succeeded in a real environment were unknown, and urged organizations to investigate possible breaches back to at least October 17. JPCERT/CC said ongoing unauthorized access in Japan might involve searching for and exploiting known vulnerabilities or management weaknesses that vary by target, and urged organizations to inspect APIs and administrative functions as well as the two PHP flaws. The PHP Foundation acknowledged that its past delays were a structural failure caused by decision-making and resource shortages and is reviewing its decision process. It said experienced engineers with knowledge of PHP as a whole are still needed to decide on fixes and review them, even as AI-generated reports increase, and called on companies using PHP and government agencies to provide sustained financial support.

PHP maintainers released fixes on October 21, 2026, for two critical vulnerabilities that could let attackers run code on servers without authentication. OpenAI said the same day that security company Calif found one of the flaws using its AI model GPT-5.6 Sol. In Japan, business, childcare and government services paused operations as a precaution, and Prime Minister Sanae Takaichi urged organizations to prepare to update PHP.

The issue in PHP itself, called “PHP2Shell,” could leave internal state inconsistent while handling memory exhaustion or fatal errors, potentially enabling remote code execution (running code on a remote server) under certain conditions. WordPress said no update to WordPress itself was needed and asked users to check their hosting providers’ guidance and update PHP. Calif’s discovery was part of OpenAI’s “Patch the Planet” initiative to support open-source maintenance. Wiz also independently found and reported vulnerabilities, saying AI helped test and debug attacks that combined multiple issues and were adjusted to different memory layouts on each target.

PHP maintainers had identified signs related to the root cause in 2024, but response stalled because no specific attack path had been found, fixes posed compatibility challenges, and staffing was limited. Reports in summer 2026 clarified the impact, making a fix the top priority. PHP maintainers led the fixes and release, with research teams and security companies helping validate them and prepare defenses.

According to the PHP Foundation, vulnerability identifiers and other details leaked online at around midnight Japan time on October 21, moving the planned release from October 22 to October 21. At around 5 a.m. that day, Calif mistakenly published a proof-of-concept video demonstrating a working attack; copies continued to circulate on social media after it was deleted. The foundation said the video accelerated development of proof-of-concept code. At 4:41 p.m., software distribution company Chainguard published images containing the fix more than six hours before the coordinated release was scheduled. The foundation said attackers could then analyze the fix.

In Japan, Chatwork carried out precautionary maintenance from the evening of October 21 into the early hours of October 22, while childcare and education services provider Codmon halted all services. Codmon advised facilities to use alternatives such as attendance and departure logs, and said in its October 21 notice that it had found no unauthorized access or information leaks linked to the vulnerability. Some Japan Meteorological Agency weather-data pages became inaccessible on the morning of October 22, though Kikikuru, typhoon information and rain-cloud forecast pages remained available, FNN reported. On the evening of October 21, Takaichi asked operators to check whether their public websites used PHP and to prepare to apply the fix scheduled for release at 11 p.m. that day.

Amazon Threat Intelligence said on October 21 that records from MadPot, a system that attracts and observes attacks, showed an exploitation attempt dated October 17. It said the route by which the attack information was obtained and whether the attempt succeeded in a real environment were unknown, and urged organizations to investigate possible breaches back to at least October 17. JPCERT/CC said ongoing unauthorized access in Japan might involve searching for and exploiting known vulnerabilities or management weaknesses that vary by target, and urged organizations to inspect APIs and administrative functions as well as the two PHP flaws. The PHP Foundation acknowledged that its past delays were a structural failure caused by decision-making and resource shortages and is reviewing its decision process. It said experienced engineers with knowledge of PHP as a whole are still needed to decide on fixes and review them, even as AI-generated reports increase, and called on companies using PHP and government agencies to provide sustained financial support.