OpenAI models help uncover two critical PHP flaws before attackers

Key Points

1. Two critical vulnerabilities were found in PHP, the language running most of the web.
2. A coalition of maintainers, researchers and security firms patched them over roughly two months.
3. OpenAI says frontier models reached the flaws first, creating a defender's window.

The latest:

Two critical vulnerabilities in PHP — the programming language behind most websites and web applications — were fixed after a roughly two-month coordinated response led by OpenAI alongside open-source maintainers, researchers and security companies. OpenAI said its frontier models reached the flaws before attackers did, and described the PHP issues as among the most consequential security problems in recent memory.

Details:

• The discovery: Security firm Calif found one of the critical vulnerabilities using the GPT-5.6 Sol model through OpenAI’s Patch the Planet initiative, according to OpenAI. Separately and independently, Wiz identified both vulnerabilities and reported them, giving the response two distinct discovery tracks rather than a single finder.
• Who fixed it: PHP’s own maintainers led the remediation and release process, OpenAI said. Wiz, Cloudflare and other parties contributed research, technical support and coordination on protective measures that were put in place before the vulnerabilities were disclosed publicly.
• The tooling: OpenAI released a PHP version-scanning tool packaged as a skill inside Codex, intended to help operators determine whether the builds they are running are exposed. The company did not set out an adoption target or a deadline for operators to complete checks.
• Defender’s window: OpenAI uses the term defender’s window for the period between a model-assisted discovery and the moment attackers could exploit the same flaw. The company’s argument is that this interval is what allows patches, mitigations and guidance to be staged in advance.
• Five-stage framework: OpenAI proposed a five-stage response model for critical flaws in widely deployed software: verify and fix the original vulnerability, including adversarial testing with frontier models; give affected organizations advance time for temporary protection; coordinate with government partners; update affected products; then publish guidance and support deployment.
• Government partners: In the PHP case, coordination included the White House Super Intelligence Force, the Office of the National Cyber Director, the Cybersecurity and Infrastructure Security Agency, and international partners, according to OpenAI. The company did not name the international partners involved.
• The funding: OpenAI announced $1 billion in subsidized access credits for its Daybreak product through a program called Daybreak for Frontline Defenders, aimed at organizations protecting essential services. It also said open-source maintainers will receive free Codex subscriptions, with support extended to projects including the Rust Foundation.
• The warning: OpenAI cautioned that the same capability that surfaces vulnerabilities can help attackers assemble offensive exploit tooling quickly. The company said it expects a sharp increase in similar discoveries, which would place repeated strain on patching cycles across the open-source ecosystem.

Background:

PHP underpins the majority of websites and web applications online, so a flaw in it propagates far beyond any single vendor. Patching depends on independent operators updating their own installations, which is why pre-disclosure coordination and scanning tools carry weight.

Between the lines:

The $1 billion in Daybreak credits and free Codex subscriptions for maintainers point to the gap the framework exposes: discovery is now faster than remediation capacity, and the volunteers maintaining critical open-source code are the bottleneck. OpenAI’s own warning that attackers can use the same capability makes that asymmetry the core risk, not the two PHP flaws themselves.

What’s next

Watch how quickly operators running exposed PHP builds apply the fixes, whether the five-stage framework is adopted by other vendors or agencies, and whether the predicted surge in model-assisted vulnerability discoveries materializes.