OpenAI Coordinates Response to PHP Vulnerabilities

Over two months, OpenAI led maintainers, researchers, and security providers in responding to critical vulnerabilities in PHP, an open-source language used for websites and web applications. OpenAI says Calif found one flaw with GPT-5.6 Sol through its Patch the Planet initiative; Wiz independently found and reported both vulnerabilities. PHP maintainers led the fixes and release, with OpenAI, Wiz, Cloudflare, and others contributing research, technical support, and coordination before disclosure. OpenAI urged PHP users to apply the fixes and shared a Codex skill that checks PHP versions so operators can identify potentially vulnerable installations and determine applicable updates using PHP’s and providers’ guidance.

OpenAI called the flaws among the most impactful security issues in recent memory, citing their severity and PHP’s widespread use. The company said frontier AI found them before attackers did, creating what it calls a “Defender’s Window” for maintainers and security providers to prepare protections before disclosure. It also warned that AI can help attackers develop offensive tools and exploits. OpenAI committed $1 billion in subsidized Daybreak access through Daybreak for Frontline Defenders, aimed at organizations with limited security resources that protect essential services. OpenAI said these bugs are among several significant vulnerabilities it has found and is working to fix, and expects similar findings to increase sharply in the coming months. It proposed a response framework for widely used software and invited the security community to improve it.

The framework has five stages: validate the upstream vulnerability and test a fix, including adversarial tests with frontier models; give affected organizations time to prepare; coordinate with government partners to protect critical systems; prepare updates across affected products; and publish guidance and support deployment once the patch is available. Maintainers decide whom to notify before disclosure. The proposed criteria include each organization’s protective role, time to act, and ability to handle sensitive information. Participants would receive information needed for their work, from network logs with indicators of compromise to more sensitive proof-of-concept exploits. The framework weighs the protection such sharing enables against the risk that even limited details reveal a flaw. Teams would reassess timelines if exploitation is active, information leaks, a public patch reveals the vulnerability, or delays increase risks to users.

The Patch the Planet website gathers public disclosures and guidance. For PHP, it links to maintainers’ patch and upgrade instructions and partner research, and covers interim mitigations, when to consult a provider or expert, and backup and recovery. OpenAI said interim protections can reduce risk but have limits and operational tradeoffs. It pointed operators to CISA’s prioritization and remediation guidance, which can help determine whether to patch urgently, mitigate immediately, or plan an update, with forensic checks where needed to assess possible compromise.

OpenAI said maintainers need funding, model access, and engineering help to sort new security reports from duplicates, assess impact, and test fixes across supported releases. Its support includes Codex for OSS subscriptions for maintainers, sponsorship of projects such as the Rust Foundation, and hands-on help through Patch the Planet. OpenAI is also calibrating additional Daybreak blocking classifiers to reduce misuse, while acknowledging safeguards can block legitimate security work. It said broad access for defenders and protections against misuse both matter.

OpenAI also outlined ways AI could help engineers build safer software: safer coding defaults and patterns, migration of critical components to memory-safe languages, formal verification (proof that software meets stated security properties), and stronger sandboxing (isolating software to limit harm if compromised). It invited maintainers, researchers, response teams, providers, and operators to comment on disclosure windows, emergency patching, early notice, and secure-by-default packages, including fixes for downstream and older versions. OpenAI also sought independent experts to review the approach and identify gaps.

Over two months, OpenAI led maintainers, researchers, and security providers in responding to critical vulnerabilities in PHP, an open-source language used for websites and web applications. OpenAI says Calif found one flaw with GPT-5.6 Sol through its Patch the Planet initiative; Wiz independently found and reported both vulnerabilities. PHP maintainers led the fixes and release, with OpenAI, Wiz, Cloudflare, and others contributing research, technical support, and coordination before disclosure. OpenAI urged PHP users to apply the fixes and shared a Codex skill that checks PHP versions so operators can identify potentially vulnerable installations and determine applicable updates using PHP’s and providers’ guidance.

OpenAI called the flaws among the most impactful security issues in recent memory, citing their severity and PHP’s widespread use. The company said frontier AI found them before attackers did, creating what it calls a “Defender’s Window” for maintainers and security providers to prepare protections before disclosure. It also warned that AI can help attackers develop offensive tools and exploits. OpenAI committed $1 billion in subsidized Daybreak access through Daybreak for Frontline Defenders, aimed at organizations with limited security resources that protect essential services. OpenAI said these bugs are among several significant vulnerabilities it has found and is working to fix, and expects similar findings to increase sharply in the coming months. It proposed a response framework for widely used software and invited the security community to improve it.

The framework has five stages: validate the upstream vulnerability and test a fix, including adversarial tests with frontier models; give affected organizations time to prepare; coordinate with government partners to protect critical systems; prepare updates across affected products; and publish guidance and support deployment once the patch is available. Maintainers decide whom to notify before disclosure. The proposed criteria include each organization’s protective role, time to act, and ability to handle sensitive information. Participants would receive information needed for their work, from network logs with indicators of compromise to more sensitive proof-of-concept exploits. The framework weighs the protection such sharing enables against the risk that even limited details reveal a flaw. Teams would reassess timelines if exploitation is active, information leaks, a public patch reveals the vulnerability, or delays increase risks to users.

The Patch the Planet website gathers public disclosures and guidance. For PHP, it links to maintainers’ patch and upgrade instructions and partner research, and covers interim mitigations, when to consult a provider or expert, and backup and recovery. OpenAI said interim protections can reduce risk but have limits and operational tradeoffs. It pointed operators to CISA’s prioritization and remediation guidance, which can help determine whether to patch urgently, mitigate immediately, or plan an update, with forensic checks where needed to assess possible compromise.

OpenAI said maintainers need funding, model access, and engineering help to sort new security reports from duplicates, assess impact, and test fixes across supported releases. Its support includes Codex for OSS subscriptions for maintainers, sponsorship of projects such as the Rust Foundation, and hands-on help through Patch the Planet. OpenAI is also calibrating additional Daybreak blocking classifiers to reduce misuse, while acknowledging safeguards can block legitimate security work. It said broad access for defenders and protections against misuse both matter.

OpenAI also outlined ways AI could help engineers build safer software: safer coding defaults and patterns, migration of critical components to memory-safe languages, formal verification (proof that software meets stated security properties), and stronger sandboxing (isolating software to limit harm if compromised). It invited maintainers, researchers, response teams, providers, and operators to comment on disclosure windows, emergency patching, early notice, and secure-by-default packages, including fixes for downstream and older versions. OpenAI also sought independent experts to review the approach and identify gaps.