OpenAI says AI helped find critical PHP flaws; fix released, $1 billion credits pledged
In short
OpenAI said on 21 October that it worked with PHP maintainers, security firms and agencies for about two months to fix two critical PHP vulnerabilities, one found by the firm Calif using GPT-5.6 Sol. PHP released fixed versions on 21 October. OpenAI also pledged $1 billion in subsidised access credits under its Daybreak for Frontline Defenders programme.
Read the full story 2 min read
OpenAI said on 21 October that over the past two months it had worked with open-source maintainers, researchers and security providers to help fix serious vulnerabilities in PHP. CLS reported OpenAI's statement as saying that, given the severity and PHP's wide use, this is one of the most impactful recent vulnerabilities. [ 1 , 3 ]
Security firm Calif used GPT-5.6 Sol to find one flaw, and security firm Wiz independently found and reported related vulnerabilities, cnBeta, Jiemian and IT Home reported. The PHP maintainers led verification, repair and release of the fixed versions, while OpenAI, Wiz, Cloudflare and other partners provided research, engineering support and coordination. IT Home said the two flaws, collectively called PHP2Shell, carry the numbers CVE-2026-109618 and CVE-2026-109619, each with a CVSS v4.0 score of 9.2. [ 2 , 3 , 4 ]
cnBeta and IT Home said both flaws were rated critical and could let attackers run code remotely without authentication. IT Home said the PHP Foundation issued an emergency update on 21 October, with fixed versions PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. IT Home and Huxiu both said about 70% of websites rely on PHP. [ 2 , 4 , 5 ]
OpenAI said that, as far as it knew, the flaw was found with the help of advanced AI before attackers found and exploited it, giving maintainers what it called a Defender's Window. It also said the same capabilities could help attackers build exploits faster, and that this window may narrow as models improve. [ 2 , 4 ]
OpenAI proposed a five-stage framework for serious vulnerabilities: verifying and fixing the flaw upstream, notifying affected organisations early, working with government agencies, preparing updates across products, and supporting deployment after the patch. cnBeta said that for the PHP case OpenAI contacted several US government bodies, including CISA, and coordinated with international partners. [ 2 , 3 , 4 ]
OpenAI pledged $1 billion in subsidised access credits through its Daybreak for Frontline Defenders programme, aimed at organisations with limited security resources that run critical services. It also said it will provide a PHP version-check tool and a Codex skill, and gather disclosure information and guidance through its Patch the Planet project. [ 1 , 2 , 3 , 4 ]
Huxiu gave a fuller account. It reported that the PHP update was brought forward from 22 October to the evening of 21 October, that several government and company services in the Netherlands, Finland and Japan went offline or into maintenance, and that Calif leaked an attack demonstration video. It also said the PHP Foundation's review admitted structural failures caused by shortages in staff and in its security team's decision process. These points come from Huxiu alone. [ 5 ]
Why it matters
PHP is described by IT Home and Huxiu as used by about 70% of websites, so a flaw allowing remote code execution without authentication could reach many sites and services. OpenAI said the case shows AI can help defenders find flaws before attackers do, while the same capability could also speed up attacks.
Key facts
- OpenAI issued a statement on 21 October about serious vulnerabilities in PHP, calling it one of the most impactful recent vulnerabilities given its severity and PHP's wide use. [ 1 , 4 ]
- Security firm Calif used GPT-5.6 Sol to find one flaw, and security firm Wiz independently found and reported related vulnerabilities. [ 2 , 3 , 4 ]
- Two vulnerabilities were rated critical and could let attackers execute code remotely without authentication. [ 2 , 4 ]
- The PHP team released fixed versions on 21 October and urged users to upgrade; IT Home listed PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. [ 2 , 4 , 5 ]
- OpenAI pledged $1 billion in subsidised access credits through its Daybreak for Frontline Defenders programme. [ 1 , 2 , 3 , 4 ]
- OpenAI proposed a five-stage response framework for serious vulnerabilities. [ 2 , 3 , 4 ]
- OpenAI said it will provide a PHP version-check tool and a companion Codex skill to help operators identify at-risk PHP versions. [ 2 , 4 ]
Confirmed by several sources
- OpenAI said that, with open-source maintainers, researchers and security providers, it helped fix serious PHP vulnerabilities, and that Calif found one flaw with GPT-5.6 Sol while Wiz independently found and reported related flaws. [ 2 , 3 , 4 ]
- PHP released a fix on 21 October for two critical vulnerabilities that allow remote code execution without authentication. [ 2 , 4 , 5 ]
- OpenAI committed $1 billion in subsidised Daybreak access for defenders through the Daybreak for Frontline Defenders programme. [ 1 , 2 , 3 , 4 ]
- OpenAI described a staged vulnerability-response framework covering verification and repair, advance notice to affected organisations, government coordination, product updates and fix guidance. [ 2 , 3 , 4 ]
- IT Home and Huxiu both said about 70% of websites rely on PHP. [ 4 , 5 ]
Still unclear
- How many flaws each party found. cnBeta and IT Home describe Calif finding one flaw and Wiz separately reporting related ones; the documents do not give a clear count for each.
- Huxiu's account of disruption on 21 October, including the Dutch digital identity system, parliament website and national cybersecurity centre website going offline, and services shut by a Finnish government IT provider and BambooHR. Only Huxiu reports this; the other documents do not mention outages.
- Huxiu's report that Calif leaked an attack demonstration video and that PHP brought its update forward from 22 October to 21 October. Single-source claim; OpenAI's statement as relayed by other outlets does not mention it.
- Whether the flaws were exploited before the fix. OpenAI said that, as far as it knows, the flaw was found with AI assistance before attackers found and exploited it; Huxiu reports services taken offline as a precaution, and no document confirms attacks.
What local media are saying
Timeline, local time
- Huxiu reported the PHP security update went live on the evening of 21 October. [ 5 ]
- CLS reported OpenAI's statement on the PHP vulnerability and the $1 billion pledge. [ 1 ]
- cnBeta published its report on the OpenAI announcement. [ 2 ]
- Jiemian reported OpenAI's phased response framework and the $1 billion Daybreak credits. [ 3 ]
- IT Home published its report on 22 October, including the fixed PHP versions. [ 4 ]
- Huxiu published its account of the discovery and aftermath. [ 5 ]