On October 21 OpenAI published a post on "Patch the Planet", an OpenAI-led effort that connects security researchers, maintainers and incident responders to deal with critical vulnerabilities in open-source software. The main example is two PHP vulnerabilities.

The idea is simple. AI can now find bugs before attackers do, but finding is only the start. Patches must be adapted for older releases, cloud providers need mitigations ready, and many organizations have to update. So downstream parties get prepared before anything goes public. OpenAI calls that stretch the "Defender's Window".

The post is frank about the downside: the same capabilities help attackers build exploits faster, and OpenAI expects findings like these to rise sharply over the coming months. How long disclosure windows should be, and when governments may patch systems they manage without consent, are listed as unanswered. Everything below comes from that one post; I found no independent verification.

Patch the Planet is not a tool. It is a coordination approach plus a website that points to maintainers' official patches, partner research, interim mitigations, and backup and recovery guidance.

It targets a specific mismatch: finding vulnerabilities is getting fast, but getting fixes into real systems is still slow. A fix is more than a code change. It may mean porting a patch to older releases, getting providers to prepare protections first, and getting many organizations to update their systems.

Over the past two months OpenAI led a coalition of maintainers, researchers and security providers on critical flaws in PHP, the language behind a lot of websites and web apps.

• Calif found one flaw using GPT-5.6 Sol, as part of the initiative.
• Wiz independently discovered and reported both vulnerabilities.
• PHP's maintainers led the fix and the release. Partners such as Wiz and Cloudflare contributed research, technical support and coordination so protections were ready before disclosure.

OpenAI says these bugs are among several significant vulnerabilities it is already working on, so more will follow.

The post breaks the process into five stages:

1. Validate and fix upstream. Maintainers, researchers and technical partners confirm the issue, assess impact and test a fix, including adversarial testing with frontier models for bypasses and variants.
2. Give affected organizations time to act. Maintainers decide whom to brief in advance. Distributors can backport fixes, hosts can test mitigations, security providers can prepare detection or WAF rules.
3. Coordinate with government partners. For PHP, OpenAI engaged several U.S. bodies, including the White House's Super Intelligence Force, the Office of the National Cyber Director (ONCD) and CISA, plus international partners.
4. Prepare updates across products. Once fixes and disclosure plans stabilize, more distributors, providers and operators come in.
5. Publish guidance and support deployment. Public guidance helps people check exposure, find updates or interim mitigations, and verify the fix is running.

On who is told first, the post names three considerations: the organization's protective role, its ability to act within the available time, and its ability to handle sensitive information. What is shared can be indicators of compromise (IoCs) and, for trusted parties, more sensitive proof-of-concept exploits (PoCs).

The hard part is that even limited details can reveal where a vulnerability is. So OpenAI weighs the protection gained from early notice against the risk of a leak. The timeline is reassessed on active exploitation, a leak, a patch that reveals too much, or growing user risk from stalled remediation. The post gives no fixed day counts.

• A PHP version checker, shared as a Codex skill, to help operators spot potentially vulnerable versions.
• The Patch the Planet website: official patches, partner research, interim mitigations, backup and recovery guidance.
• CISA BOD 26-04: guidance on prioritizing remediation by exposure, exploitation and impact.
• $1 billion in subsidized Daybreak access through Daybreak for Frontline Defenders, aimed at resource-limited organizations protecting essential services.
• Codex for OSS subscriptions for all maintainers, direct sponsorship of projects such as the Rust Foundation, and application-based hands-on support run with Trail of Bits.

OpenAI also says it will help engineers write safer code, migrate critical components to memory-safe languages, apply formal verification to widely used components, and strengthen sandboxing for software that processes untrusted files.

• How long should disclosure windows be?
• When may governments patch systems they manage without consent?
• How do you give early notice without raising leak risk?
• Blocking classifiers for Daybreak are still being calibrated. OpenAI admits safeguards can block legitimate security work and calls the balance difficult.
• Interim protections reduce risk but have limits and operational trade-offs; OpenAI says it will state them clearly.

OpenAI asks the security community for feedback and for independent experts to review and challenge the framework.

What follows is my own read, not the post's words.

First, OpenAI wrote this itself. Who found and who fixed what is its account, and I found nobody who checked it. Read it as one company's published approach, not an industry standard.

Second, the practical point for ordinary developers: the projects you depend on are more likely to get AI-found vulnerabilities, and the news may circulate in trusted channels before it is public. If your upgrade routine runs once a week, you will miss the window. The actions are concrete: watch official security advisories for your languages and frameworks, and keep a fast path in your deploy pipeline for shipping patch versions.

Third, the post sets no fixed number of days. That is honest, but it also means you cannot treat it as an SLA.

coordinated disclosure

After finding a vulnerability, let maintainers and downstream parties prepare first, then go public.

Defender's Window

OpenAI's term for the preparation time defenders get when AI finds a bug before attackers do.

backport

Porting a fix from a newer version to older ones.

WAF

Web application firewall; rules can block known attack requests.

IoC / PoC

Indicator of compromise / proof-of-concept exploit.

Daybreak

OpenAI's cyber-defense program for defenders; the post mentions subsidized access and blocking classifiers. I did not look into it further.

• OpenAI: Patch the Planet — global response coordination as agents uncover critical vulnerabilities (2026-10-21)

I read this page through a summarizing tool, not word for word. Defer to the original for every number and name.