AI-assisted discovery of critical PHP flaws prompts coordinated fixes

OpenAI says it led a coalition for two months to address critical vulnerabilities in PHP, an open-source language used to run websites and web applications.[2] The effort drew on security researchers from start-ups and major technology companies, as well as U.S. and international government officials.[1] OpenAI says PHP maintainers led the fixes and release, while OpenAI, Wiz, Cloudflare and others contributed research, technical support and coordination.[2] That division of work matters: finding a flaw is only one step, and the maintainers responsible for the underlying software led the release of fixes.[2]

Wiz says the vulnerabilities included memory disclosure and unauthenticated remote code execution in PHP itself, independent of applications built on the language.[3] The company says PHP powers 70% of websites worldwide and that every PHP version released in the previous decade was affected by at least one vulnerability, reaching back to version 5.4.[3] Wiz grouped the flaws under two CVEs and reported a severity score of CVSS 9.2.[3] PHP released a patch for multiple memory-safety vulnerabilities in its core engine on October 21, 2026.[3] Teams should assess affected PHP versions and prioritize patch planning.

OpenAI says its Calif system found a critical PHP flaw using GPT-5.6 Sol, while Wiz independently discovered and reported both vulnerabilities.[2] OpenAI characterized the PHP flaws as among the most impactful security issues in recent memory, citing their severity and PHP’s widespread use.[2] Fouad Matin, who leads OpenAI’s cyber-preparedness and safety efforts, said he was running on about two hours of sleep while working to patch two critical vulnerabilities discovered by AI.[1] He also said OpenAI had several cybersecurity teams and was devoting significant computing power to finding and fixing problems.[1]

OpenAI expects similar findings to increase sharply in the coming months.[2] Daybreak had developed and tested 661 patches, with 458 accepted by maintainers.[4] Cybersecurity researchers were advised to submit high-quality, readable reports and devote resources to coordinating disclosure.[4] In other AI security disclosures, Anthropic said it detailed five attack chains with critical impacts requiring little or no user interaction.[5] Its disclosure process gives developers 90 days to patch, followed by a further 45 days before additional details are discussed.[5] Anthropic also said Google had assigned 62 CVEs to its Chrome findings and patched them as of Chrome 156.[5]