OpenAI recently announced its involvement in discovering and coordinating the fix for two critical security vulnerabilities in the PHP programming language. This initiative is driven by OpenAI's "Patch the Internet" program, which brings together software maintainers, security researchers, cloud service providers, and government cybersecurity agencies to collectively mitigate the risks posed by these vulnerabilities to global internet infrastructure. OpenAI stated that advanced AI models are now capable of identifying software security flaws with wide-ranging impacts, and how to rapidly translate these findings into practical protective measures is becoming an increasingly important issue in the field of cybersecurity.
In this incident, the security company Calif utilized GPT-5.6 Sol to discover a serious security flaw, while the security company Wiz independently identified and reported two related vulnerabilities. The PHP maintenance team led the efforts to fix the vulnerabilities and release secure versions, with OpenAI, Wiz, Cloudflare, and other partners providing technical research, engineering support, and coordination to help relevant organizations prepare defenses before the vulnerabilities were disclosed. PHP is an open-source programming language widely used for websites and web applications, with numerous internet services relying on it to handle web requests, database operations, and backend business logic.
As a result, the impact of serious vulnerabilities at the PHP core level could quickly extend to numerous websites, cloud services, and enterprise systems.
OpenAI Assists in Fixing PHP Vulnerabilities Through AI Technology
The two vulnerabilities discovered were both rated as critical, potentially allowing attackers to execute code remotely without authentication. The PHP development team released a patch on October 21 and urged users to upgrade immediately. OpenAI stated that this incident highlights the dual role of artificial intelligence in cybersecurity. On one hand, advanced models can help researchers uncover long-hidden flaws in mature software; on the other hand, the same capabilities could be exploited by attackers to rapidly develop exploit tools and attack programs.
Therefore, security teams cannot focus solely on whether vulnerabilities can be discovered; they must also establish an efficient mechanism to ensure vulnerabilities are promptly fixed and to minimize attack risks as much as possible before relevant patches are deployed.
According to current information, the vulnerabilities were discovered with the assistance of advanced artificial intelligence before attackers could exploit them, providing software maintainers with valuable time to prepare defenses. OpenAI refers to this phase as the "defender's window," a time frame in which defenders can identify issues, formulate fixes, and deploy protective measures ahead of attackers. However, this window will not last indefinitely. As the capabilities of AI models continue to improve, the time lag between discovering vulnerabilities and developing attack methods may gradually shorten.
To address this, OpenAI proposed a coordinated response framework for critical vulnerabilities, which primarily includes five stages. First, software maintainers, researchers, and key technical partners jointly verify the vulnerabilities, assess their impact, develop and test fixes, and conduct adversarial testing using advanced models to check whether the patches effectively prevent attacks and to identify related vulnerability variants. Second, sufficient preparation time is allocated for affected organizations, providing necessary technical information in phases according to the actual protective responsibilities of each participant. Third, collaboration with government cybersecurity agencies ensures that critical infrastructure operators receive timely alerts.
In the PHP vulnerability incident, OpenAI and its partners contacted the U.S. White House Super Intelligence Task Force, the Office of the National Cyber Director, and the Cybersecurity and Infrastructure Security Agency (CISA), and coordinated with international partners.
Fourth, once the fix and public disclosure plans stabilize, the scope of participation is expanded to allow more software publishers, cloud service providers, hosting service providers, and enterprise security teams to prepare updates in advance. Fifth, after the patch is released, clear remediation guidelines are provided to the public, and deployment status is continuously tracked. OpenAI specifically noted that sharing vulnerability information in advance presents a clear dilemma. Cloud service providers, cybersecurity vendors, and critical infrastructure operators need sufficient technical details to effectively prepare protective measures, but even limited vulnerability information could help attackers infer the location of flaws.
Therefore, relevant information should be provided in phases based on the protective responsibilities of participants, preparation time requirements, and their capacity to handle sensitive information. If vulnerabilities have been actively exploited, technical details inadvertently leaked, or the public patch itself exposes the location of vulnerabilities, the original disclosure timeline should be reassessed. To help more organizations take action, OpenAI stated it will provide a PHP version checking tool and plans to launch accompanying Codex skills to assist system operators in identifying potentially vulnerable PHP versions. OpenAI is also compiling vulnerability disclosure information and protective guidelines through the "Patch the Internet" project.
In terms of resource support, OpenAI has committed to providing $1 billion in subsidized Daybreak service access through its "Dawn of Frontline Defenders" program, focusing on assisting organizations with limited security resources that undertake critical service protection tasks; offering subscription support to open-source software maintainers through Codex for OSS; and directly sponsoring projects like the Rust Foundation. OpenAI also suggested that security flaws should be reduced from the software design phase, including helping engineers choose safer default settings, migrating critical components to memory-safe languages, and adopting formal verification and stronger isolation mechanisms.
The company acknowledged that expanding the use of AI in cybersecurity requires more robust anti-abuse measures and is adjusting security classifiers for Daybreak to strike a balance between open defensive capabilities and preventing malicious use.
Finally, OpenAI called on software maintainers, researchers, cybersecurity service providers, and enterprise operators to collaboratively refine this framework and discuss issues such as preparation time before vulnerability disclosure, conditions for early disclosure, alert recipients, and secure alternatives for outdated software. The company also proposed a more controversial future vision: internet systems could gradually establish continuous vulnerability scanning mechanisms to notify website owners upon discovering known vulnerabilities and even directly patch systems with authorization, but this raises concerns about system ownership, administrative permissions, risks of misoperation, and liability boundaries, which still require strict technical and governance safeguards.