OpenAI's Daybreak gave vetted defenders a head start on critical PHP flaws

OpenAI says its strongest cyber models, handed to vetted defenders, found critical PHP vulnerabilities before attackers did. According to Cloudflare's CEO, exploit attempts rose from thousands before the announcement to millions afterward.

That is the shape of the story: a genuine head start, a genuine race, and a company blog post that is more confident about the first than the evidence in it can support.

What OpenAI says happened

OpenAI says it spent two months leading a coalition of maintainers, researchers and security providers to address critical vulnerabilities in PHP, "an open-source programming language used to run websites and web applications." Axios, reporting on October 21, puts the scale in plain terms: PHP is used on nearly 70% of all websites.

The discovery credit is shared. A security researcher at Calif found a critical flaw using GPT-5.6 Sol as part of OpenAI's Patch the Planet initiative, and Wiz independently discovered and reported both vulnerabilities, according to OpenAI. PHP's maintainers led the fixes and the release. OpenAI, Wiz, Cloudflare and others contributed research, technical support and coordination ahead of disclosure.

Then comes the sentence doing the heavy lifting: "To our knowledge, frontier AI found these vulnerabilities before attackers did, opening a Defender's Window for maintainers and security providers to prepare fixes and protections ahead of disclosure." Note the opening clause. "To our knowledge" is a footnote wearing a sentence.

The machinery behind it

The vehicle is Daybreak, which OpenAI describes as "a governed cyber defense stack": frontier models, the Codex harness, Codex Security, trusted workflows and ecosystem partners. It runs what OpenAI calls an agentic defense loop of six stages: inventory, discover, validate, assign, remediate and prove.

The access side is Daybreak Access. Verified defenders get "more capable and permissive defensive tools paired with stronger verification, scope controls, and oversight." That is the policy bet in one line: loosen the models for the people you have checked, and keep a hand on the leash.

What the materials do not include is the vetting bar. OpenAI's cited materials do not spell out the qualification criteria or what "permissive" means in practice. For a program whose whole premise is that these models are dangerous in the wrong hands, that is the table I would most like to read.

Cloudflare offers a concrete product example. Its early-access Vulnerability Discovery and Remediation service, offered through Cloudflare Managed Defense, uses OpenAI models accessed through Daybreak to examine customer-authorized code and check the evidence behind each finding. It then prepares proposed patches and, where appropriate, custom web application firewall rules to reduce exposure while developers review a fix.

OpenAI is also distributing a PHP version checker as a Codex skill, so operators can identify potentially vulnerable PHP versions and work out which updates apply.

The numbers after disclosure

The figures after the announcement are the sobering part. Cloudflare CEO Matthew Prince, per Axios, put numbers on the surge: "That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now." Government officials from the U.S. to Japan are warning companies to patch or consider taking their sites down, Axios reports.

Wiz's Ami Luttwak called it, in remarks to the Wall Street Journal quoted by The Verge, "the most severe AI vulnerability that was ever found." That is a superlative from a firm that also discovered the bugs, so weigh it accordingly. It is not a measurement.

OpenAI describes a window for maintainers and providers to prepare protections before disclosure. Prince's reported surge in attempts shows why operators still face pressure to update. Preparation is not the same thing as deployment.

What is still unclear

The coverage does not line up neatly. Axios says OpenAI's model contributed to finding one of the vulnerabilities. The Forbes piece, as summarized by Techmeme, says Project Daybreak helped uncover PHP2Shell RCE bugs, plural. OpenAI's own account credits Calif's GPT-5.6 Sol work with one critical flaw and Wiz with independently reporting both. The Verge also calls Wiz "Google's Wiz security team," while OpenAI's materials describe it only as a partner and security provider.

The open items, then:

• The CVE identifiers and affected PHP version numbers are not given in the coverage cited here.
• Whether Wiz used AI models or traditional methods to find both flaws is not stated.
• Whether anyone exploited the bugs before patches and firewall rules were available is not stated. OpenAI's claim is limited to its own knowledge.

What to watch

Watch for three things: the CVE details, which would let outsiders check the severity claims; any public account of who gets Daybreak Access and why; and whether the exploit-attempt numbers Prince described fall as patching catches up.

OpenAI says defenders got preparation time. Prince reports what followed disclosure: millions of attempts. The missing measure is how much exposure those preparations prevented. Show me the eval, not the vibe.