OpenAI Says AI Found the PHP Flaws Before Attackers. Cloudflare Says They Leaked and Drew Attack Traffic

OpenAI said on October 21 that, to its knowledge, frontier AI found the two critical PHP flaws before attackers did, while Cloudflare said the same day that the flaws leaked before the official release and that it saw attack traffic against its customers. Both accounts concern the vulnerabilities tracked as CVE-2026-109618 and CVE-2026-109619.

The first account is OpenAI's, as reported by PPC Land. OpenAI's own page returned an access error when fetched, so its wording here is quoted through that report. The second is Cloudflare's own blog post.

What OpenAI claimed about the PHP flaws

OpenAI says Calif used its GPT-5.6 Sol model to find one critical flaw under its Patch the Planet initiative. It says Wiz independently found and reported both flaws, per PPC Land.

OpenAI says it led a coalition for about two months and that PHP's maintainers led the fix and the release. Its framing, as quoted by PPC Land: "To our knowledge, frontier AI found these vulnerabilities before attackers did."

The Decoder adds that Wiz credited Gemini 4 Argon and GPT-5.6 Sol, and that the work ran across five phases involving Cloudflare, CISA and the White House.

What Cloudflare reported

Cloudflare's post gives the name php2shell and says the flaws affect servers running PHP 5.4 or later. It says Calif.io, Wiz, OpenAI and Quarkslab disclosed them.

The post says the vulnerability "was leaked before the official release date", that Cloudflare observed attack traffic, and that attackers mutated payloads to try to bypass its rules. Default WAF protection for Free and paid customers took effect at 3:02 UTC on October 21.

That is about nine hours before public disclosure, according to the post. OpenAI's account, as relayed by PPC Land, does not mention the leak or the traffic.

Both can be true. A flaw can be found by defenders first and still leak before a patch ships. The question OpenAI's claim cannot answer is how long attackers held it.

The five-stage process OpenAI proposes

PPC Land lists OpenAI's proposed stages as validate and fix upstream, give affected organisations advance notice, coordinate with government partners, prepare updates across products, and publish guidance once a patch exists. OpenAI names CISA and ONCD among the government partners.

The post gives no CVE numbers, affected versions or total finding count, per PPC Land. It also omits Quarkslab, which Cloudflare credits.

What the Patch the Planet dashboard shows

Trail of Bits runs the program with OpenAI. It launched on June 22, and Engadget reported that engineers worked with 19 projects in the first week and found hundreds of legitimate bugs, including 51 issues of which 19 were fixed.

The program page now shows 68 codebases under review and 1,995 issues found and reported. It is undated, so the counts are as seen on October 22.

1. Issues found and reported1995 issues
2. Awaiting patch1316 issues
3. Patch written679 issues
4. Patches accepted upstream466 issues
5. Patches open upstream213 issues

Source: Trail of Bits, trailofbits.com/patch-the-planet, accessed 2026-10-22

Trail of Bits puts patches accepted at 23 percent of issues found and patches written at 34 percent. The stages overlap, so the bars are not a sequence that adds up.

The firm's earlier regression hunt used OpenAI models on open-source repositories too, and the site's PHP release coverage tracked the 14:00 UTC announcement.

OpenAI expects findings of this scale to become more frequent in the coming months, per The Decoder. The OpenAI page, as summarised by PPC Land, does not say whether the PHP timeline was reassessed after the leak.