OpenAI's Models Found a 9.2-Critical Flaw in 70% of the Web. Nobody Paid for It, and the Money Went to the Edge

CVE-2026-109618 and CVE-2026-109619 are unauthenticated remote-code-execution bugs that PHP rated CVSS 9.2, one notch below the 10.0 claimed by a leaked post. They affect every PHP version back to 5.4, released in 2012, and they sit in the runtime rather than in any application, so even a three-line script on an affected version is exploitable. PHP published fixes on Oct. 21: 8.5.12, 8.4.27, 8.3.36 and 8.2.35.

None of that is new in kind; a runtime-level memory bug is as old as the language. What is new is who ran the response. Over two months a coalition of maintainers, researchers and security providers worked the problem, a security firm found a critical flaw with GPT-5.6 Sol inside a program OpenAI funds, and the lab published the global coordination account and the operator guidance itself. OpenAI did not do it alone \u2014 the PHP project led the fix, HackerOne carried triage, Wiz and Calif supplied the findings \u2014 but no other party had both the models and the balance sheet to coordinate at this scale. the lab is becoming the industry's disclosure utility, and on the evidence so far it is not being paid for it.

The coordination broke in public before it broke in private. On Oct. 20, a Hacker News post named both CVEs and claimed a 10.0 score PHP had never assigned; the post was deleted, but proof-of-concept code was circulating, and PHP's own policy reserved the right to publish early once details surfaced. It published at 13:30 UTC on Oct. 21. In the gap, UK, US, EU and Australian authorities told organisations to take systems offline if they could not patch, and the Finnish government, the Dutch national security agency and the Wikimedia Foundation took sites down. the cost of the coordination failure was paid by operators, not by the lab that coordinated it.

AI did not break the bug bounty. It broke the price of finding a bug. Anthropic says Claude Mythos Preview has found thousands of zero-day vulnerabilities across every major operating system and browser, and has extended access to more than 40 organizations that maintain critical infrastructure. China's 360 Security says its Tulongfeng model found 3,432 software vulnerabilities, 105 of them confirmed by Chinese authorities \u2014 a claim Reuters could not independently verify. Meanwhile, on the HackerOne platform the backlog of validated, unresolved findings grew 131% in two years even as average resolution time fell from 135 days to 62. the scarce input moved from finding bugs to triaging them.

Only four branches of PHP received a fix. Versions 8.1 and older are end-of-life and receive no updates at all, which leaves the 14-year tail, 5.4 onward, to hosting providers, operators and web application firewalls. CERT-EU's mitigations are explicitly partial: constrain memory, disable upload progress, switch off fatal-error backtraces. None of them makes an installation safe. machine-speed discovery does not shorten a 14-year upgrade cycle.

Patch the Planet is OpenAI's answer to a volume problem the bounty market cannot solve. Launched in June 2026 as a Daybreak initiative with Trail of Bits, it pairs frontier models with full-time human security engineers, reviews every finding before a maintainer sees it, and writes the patch with the project instead of filing a report. the lab subsidises demand, it does not sell the disclosure, and the published design carries no per-finding reward for outside hunters.

Follow the money and the structure is clear. OpenAI is subsidising the demand side with $1 billion of Daybreak access, training and support for frontline defenders to be consumed over six months, plus a defense network of more than 35 partner products. The maintainers are paid by sponsors: Automattic, JetBrains and the Sovereign Tech Agency at platinum, GoDaddy and Laravel at gold. What is left to charge for is enforcement and verification. the vendors who get paid enforce and verify, and the lab keeps the rest as marketing.

AI did not break the software, and it did not break the bug bounty \u2014 it broke the price of finding a bug. That pushes value toward the edges that enforce and the platforms that verify, and it turns the frontier lab into the industry's disclosure utility, which is a real position and, on the evidence so far, not a business. own enforcement and verification, treat discovery as a commodity, and treat the lab's security program as a funnel that has not yet decided whether to become a franchise.