OpenAI AI Model Finds Catastrophic PHP Security Flaw While Separate Audit Exposes Gaps in Its Own Monitoring System

An OpenAI model helped uncover a set of critical vulnerabilities in PHP, a programming language that underpins an estimated 70% of the world's websites and millions of applications, in one of the most consequential cybersecurity discoveries yet made with AI. The flaws, collectively called PHP2Shell, were identified by teams at OpenAI and Google-owned cybersecurity firm Wiz, among others, and could allow attackers to remotely execute code, read sensitive memory, and expose customer data. The PHP Foundation has released a patch and is urging operators to update immediately, while U.S., European, and British authorities recommended some sites go temporarily offline until fixes could be applied. Cloudflare, which helped shield websites before public disclosure, worked alongside OpenAI on the response. OpenAI president Greg Brockman and Cloudflare CEO Matthew Prince both appeared publicly to discuss the finding.

The vulnerability was discovered through OpenAI's program giving cybersecurity companies access to its ChatGPT models, which are trained to find software flaws. OpenAI said it brought the issue to PHP's attention before hackers could exploit it more severely, and described PHP2Shell as "one of the most impactful security issues in recent memory." Several government and company websites briefly went dark on October 22 as operators scrambled to apply patches.

Separately, nonprofit AI safety organizations METR and Redwood Research disclosed on October 22 that a pilot audit of OpenAI's internal agent monitoring system, conducted from July 21 to August 11, uncovered several novel vulnerabilities. The most serious exploit discovered would allow AI agents to prevent hundreds of recent actions from being recorded by the monitoring system. The evaluators had chaperoned access to OpenAI's codebase, documentation, and some technical staff during the review period. OpenAI told the evaluators it has since patched all identified vulnerabilities affecting its production safeguards, though METR and Redwood said they agreed to publish only a high-level account of their findings.