OpenAI AI Discovers Critical PHP Vulnerability: A Security Alert Affecting 70% of Websites

OpenAI, through its "Patch the Planet" initiative, used AI models to discover a critical security vulnerability in the open-source programming language PHP. PHP powers approximately 70% of websites worldwide, and OpenAI described this finding as "one of the most severe vulnerabilities in recent years." More notably, the entire process from discovery to remediation demonstrates AI's new role in cybersecurity—finding vulnerabilities before attackers do, buying time for defenders.

Vulnerability Discovery: AI Unearths a High-Risk Flaw in the PHP Engine

According to OpenAI's official disclosure, its AI models assisted researchers in discovering a critical security vulnerability in PHP. As an open-source programming language, PHP is estimated to be used by about 70% of websites, including numerous content management systems and web applications. What makes this vulnerability particularly notable is the extreme difficulty of fixing it—it requires major changes to the PHP engine.

OpenAI attributed this discovery to the "Patch the Planet" initiative, which aims to use AI models to proactively find and fix security flaws in open-source software. Compared with traditional vulnerability hunting methods, AI models can rapidly locate potential risk points across massive codebases, significantly shortening the cycle from discovery to confirmation.

Given the severity of these vulnerabilities and PHP's widespread use, we believe this is one of the most impactful vulnerabilities in recent years. It also marks a new phase: AI can discover security vulnerabilities with internet-scale impact. —OpenAI official statement

The PHP engine (typically referring to the Zend Engine) is the core execution component of the PHP language, responsible for lexical/syntax parsing, compilation to opcode, memory management, and extension interfaces. Making "major changes" at the engine level means the modifications touch underlying behavior shared by all PHP applications, posing risks far greater than fixing a specific extension or framework code: backward compatibility must be maintained while covering different version branches and numerous SAPI scenarios (such as FPM and Apache modules). As a result, such fixes often require longer review, regression testing, and release coordination cycles—which explains why the remediation needed to be prepared in advance by a coalition rather than deployed immediately.

Coordinated Remediation: The Engineering Challenges Behind a Two-Month Preparation Period

Prior to disclosure, OpenAI led a coalition of maintainers, researchers, and security vendors that spent approximately two months preparing fixes and mitigations. This timeframe was no coincidence—the fix involved core changes to the PHP engine, requiring compatibility and stability assurances to avoid introducing new issues during patching.

This "fix first, disclose later" coordinated approach is not unfamiliar in the security community, but having an AI vendor lead coordination across multiple stakeholders remains rare. OpenAI stated that all websites using PHP can now safely upgrade their versions, and can also adopt mitigation measures provided by various security vendors.

You may not have noticed that the complexity of the fix means some sites may not be able to upgrade immediately. For these scenarios, security vendors' temporary protective measures become a critical transitional solution.

The "fix first, disclose later" approach corresponds to a security practice commonly known as Coordinated Disclosure: the discovering party privately coordinates with maintainers and affected vendors before publicizing details, providing patches, mitigations, and upgrade guidance, and agreeing on a disclosure time window. Its purpose is to reduce the probability of large-scale exploitation during the "window before a patch is available." It is uncommon for an AI vendor to lead coordination across multiple stakeholders, as this requires simultaneously possessing vulnerability verification capabilities, cross-organizational communication skills, and respect for the open-source community's release cadence; poor coordination could result in information leaks or insufficient patch quality.

Defender's Window: The Time Race Between AI and Attackers

OpenAI called this incident a practical example of the "Defender's Window." The core logic is: if frontier AI can discover vulnerabilities earlier than attackers, defenders have time to prepare patches and mitigations, thereby preventing large-scale exploitation events.

Behind this concept is a realistic recognition of AI's dual-use nature—the same model capabilities can be used for both attack and defense. OpenAI claims that, to its knowledge, frontier AI discovered the vulnerability before attackers in this incident. However, this claim currently comes solely from OpenAI's unilateral disclosure, with no third-party independent verification of whether attackers indeed failed to find the vulnerability earlier.

From an industry perspective, AI-assisted vulnerability discovery is moving from the laboratory to practical application. Several research projects have previously demonstrated the potential of large language models in code auditing, but cases with genuine internet-scale impact remain rare. The discovery of the PHP vulnerability could become a landmark event in this direction.

The "Defender's Window" can be understood as the time gap in a vulnerability's lifecycle: a window exists between when a vulnerability is discovered by one party and when it is widely exploited. If defenders obtain actionable information earlier and complete patching, they can compress the window to a degree that makes large-scale exploitation difficult for attackers. This concept is closely tied to "AI dual-use": models can be used for code auditing, fuzz testing seed generation, and patch suggestions, but can also be used for reverse engineering and vulnerability localization. The point of contention lies in the evidence chain: proving that "AI discovered it before attackers" typically requires an auditable discovery timeline, independent reproduction, and third-party assessment; otherwise, it can only be regarded as a vendor's self-report.

Industry Impact: A New Variable in Open-Source Security

The PHP vulnerability incident raises new questions for the open-source security ecosystem. On one hand, improved efficiency in AI-driven vulnerability discovery means more flaws will be exposed faster, increasing remediation pressure on maintainers. On the other hand, if AI vendors can proactively coordinate fixes and disclose to the community, a new security collaboration paradigm may emerge.

For website operators using PHP, the priority is to confirm version status and upgrade as soon as possible. OpenAI emphasizes that upgrading is the fundamental way to ensure security, with security vendors' mitigations serving as a supplement.

From a longer-term perspective, this incident also sparked discussion about the transparency of AI security tools. OpenAI did not disclose specific technical details of the vulnerability in its public statement, which aligns with responsible disclosure conventions but also makes it difficult for outsiders to independently assess the actual contribution of AI in the discovery process. As more similar cases emerge, how to balance security confidentiality with public awareness will remain an ongoing challenge for the industry.

(This article is based on OpenAI's official tweet content, with some background information supplemented from industry common knowledge.)

One-line Summary

OpenAI says its AI discovered a high-risk vulnerability in PHP and coordinated the fix, demonstrating the "Defender's Window."

Paragraph Summary

Through its "Patch the Planet" initiative, OpenAI claims its AI models discovered a critical security vulnerability in PHP, which powers about 70% of websites, and described it as one of the most severe vulnerabilities in recent years. Because the fix required major changes to the PHP engine, OpenAI led a coalition of maintainers, researchers, and security vendors that spent approximately two months preparing patches and mitigations, adopting a coordinated fix-first-disclose-later model. OpenAI framed the incident as an example of the "Defender's Window": if frontier AI discovers vulnerabilities before attackers, defenders can gain time to prevent large-scale exploitation. However, the AI's actual contribution and the claim of "discovering before attackers" still lack independent third-party verification, and the vulnerability's technical details were not disclosed, making transparency and collaboration paradigms the focus of subsequent discussion.