OpenAI launched Patch the Planet with HackerOne to coordinate AI-found CVEs

Agents that read source code at the scale OpenAI runs them keep finding real bugs, and now there is a program to decide what happens next. OpenAI announced Patch the Planet on 21 October, run with HackerOne, as a channel for AI-discovered vulnerabilities to reach maintainers before they reach the open web. The Hacker News thread discussing the launch has 249 points.

The announcement is the one coordinated path behind the week's PHP story. The researcher named Calif, invited into the program on 1 August, used GPT-5.6 Sol to find the remote code execution flaw that became CVE-2026-109618 and CVE-2026-109619, published a proof-of-concept to the PHP team between 11 and 14 August, and did not post the find publicly until Cloudflare, Wiz and the PHP maintainers had patches ready to ship this week. The initial finding is the researcher's work, under the program's rules.

What Patch the Planet does

The program handles three things, according to OpenAI's post: vulnerability triage, coordinated disclosure, and focused discovery runs where researchers point agents at specific projects. HackerOne provides the reporting and payout layer. The researcher who finds a flaw works under embargo until the maintainer has fixed it, and the program sits between the agent's output and the public write-up.

OpenAI calls out that the model is a tool, not the whole method. Calif's own words from the thread are that "no model is simply capable of unearthing all bugs with simple prompts," and that the discovery still took deep domain expertise with PHP's parser. Wiz Research reached the same bug family at about the same time, using a mix of its own Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol, which is independent confirmation that the vulnerability was waiting to be found.

The question the thread raises

A Hacker News commenter wrote that OpenAI is creating a problem it then offers to solve, with the model finding bugs that would not have been found without it. The practical answer is that the bugs were there regardless, and a coordinated path to a patch is better than a public dump with exploit code attached. The subtler point is who pays for the coordination work when the next agent run finds a hundred flaws across a dependency a maintainer runs alone on evenings.

For a team shipping a product with any PHP, OpenSSL or similar long-tail code in the stack, the useful move this week is to subscribe to the maintainer's security list of each named project, so the next time Patch the Planet quietly delivers a CVE, the notice arrives before the exploit does.

Source

• Coordinated disclosure of critical vulnerabilities (OpenAI)
• Hacker News discussion