AI Turns Years-Old PHP Vulnerabilities Into Critical Threats as Millions of Exploit Attempts Hit the Web
Artificial intelligence is changing the economics of vulnerability exploitation, and the recent PHP2Shell security crisis offers one of the clearest examples yet. Critical weaknesses in PHP, some involving issues that researchers had understood for years, became the subject of an emergency global patching effort after security specialists used advanced AI models to develop practical exploitation techniques. Within hours of public disclosure, attackers launched millions of requests targeting potentially vulnerable servers, highlighting how quickly the gap between vulnerability discovery and attempted exploitation is narrowing.
The two vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, affect the PHP runtime used by a substantial portion of the internet. Their disclosure in October 2026 triggered warnings from security vendors, emergency software updates and extensive scanning activity. Cloudflare reported millions of exploitation attempts shortly after technical details became public, while Amazon Web Services identified suspicious activity targeting the flaws several days before patches were released.
Despite the scale of the response, the available evidence did not establish a corresponding wave of successful compromises. The PHP Foundation reported that it had not observed confirmed successful exploitation at the time of its statements, and researchers found that many publicly circulating exploits were ineffective. The distinction is important: the PHP2Shell incident demonstrates an acceleration in vulnerability research and attempted attacks, but it does not prove that AI has made sophisticated exploitation universally accessible or reliable.
Reporting by WIRED highlights a particularly significant aspect of the incident. AI did not merely help attackers respond to newly published vulnerabilities; it assisted legitimate security researchers in transforming previously understood weaknesses into practical attack scenarios. That development raises broader questions about how many longstanding software defects may become exploitable as AI systems improve, and whether maintainers can address them before attackers develop working techniques.
AI-Assisted Research Transforms Previously Impractical Bugs
The PHP2Shell vulnerabilities illustrate the difference between identifying a software defect and demonstrating that it can be exploited under realistic conditions. Security researchers frequently discover memory management errors, unexpected program states and other weaknesses that appear dangerous but cannot immediately be converted into reliable attacks. Some defects remain unresolved because their practical security impact is uncertain, particularly when exploitation requires complex interactions with internal software components.
Researchers at security company Calif were among the first to demonstrate practical exploitation of the PHP weaknesses in early August 2026. Their investigation combined advanced AI tools with extensive human expertise in PHP internals, memory management and vulnerability research. The work was conducted partly through OpenAI's Patch the Planet initiative, which supports the use of frontier AI systems to identify and investigate security weaknesses in widely deployed software.
According to WIRED, elements of the underlying problems had been identified as early as 2024 but were considered too difficult to exploit at the time. The later research demonstrated that assumptions about exploitability can change when investigators gain access to more capable analytical tools. Researchers at OpenAI and Wiz, Google's cloud security subsidiary, subsequently identified related PHP vulnerabilities independently using AI-assisted techniques.
The significance extends beyond the discovery of two individual bugs. Traditional vulnerability assessments often distinguish between theoretically dangerous behavior and defects that can be exploited in practice. AI-assisted analysis may narrow that distinction by helping researchers explore complex execution paths, generate test cases and investigate combinations of conditions that would otherwise require substantial manual effort.
However, the successful research did not result from an AI model independently identifying every weakness through a simple instruction. Human specialists provided the technical direction, interpreted findings and developed the expertise needed to validate exploitation. The case demonstrates how AI can amplify skilled security research rather than eliminate the need for experienced vulnerability analysts.
Millions of Exploitation Attempts Follow Emergency Disclosure
Once the PHP2Shell vulnerabilities became public, attackers moved quickly to investigate exposed systems. Cloudflare observed millions of requests associated with attempted exploitation within hours, demonstrating the scale at which internet-facing vulnerabilities can attract automated scanning and attack traffic. The activity placed additional pressure on administrators already working to identify affected PHP installations and deploy emergency updates.
The vulnerabilities are particularly concerning because they affect the language runtime rather than a specific application or plugin. PHP is widely used across content management systems, e-commerce platforms, business applications and custom web services. A weakness in its core execution environment can potentially affect multiple applications regardless of whether their own source code contains an obvious security defect.
The most serious attack scenarios involve unauthenticated remote code execution, allowing an attacker to execute instructions on a vulnerable server without first obtaining valid credentials. Depending on the deployment configuration, successful exploitation could provide access to application files, database credentials, configuration secrets and other resources available to the compromised PHP process. The resulting damage would depend on the permissions assigned to the service and the additional security controls protecting the environment.
Yet the volume of malicious requests should not be confused with the number of successful intrusions. Internet-wide exploitation statistics frequently include repeated attempts against the same hosts, automated vulnerability probes and payloads that fail because the target is patched or the required exploitation conditions are absent. In the PHP2Shell case, security researchers observed numerous attempts that did not demonstrate reliable code execution.
The PHP Foundation's statements about the absence of confirmed successful exploitation provided an important counterweight to predictions of immediate widespread compromise. Nevertheless, the discovery of attack attempts before public disclosure showed that the risk could not be evaluated solely by examining activity that began after the emergency patches were released.
Attackers Were Probing PHP2Shell Before Patches Became Available
Amazon Web Services reported identifying suspicious PHP2Shell exploitation attempts dating back to October 17, four days before the emergency security updates became available. The activity was discovered through analysis of AWS-operated honeypots, systems deliberately configured to attract malicious traffic and help researchers understand attacker behavior.
The requests originated from infrastructure associated with a commercial anonymization service, making it difficult to establish the identity or location of the operator. AWS said the activity appeared consistent with an attempt to exploit the vulnerabilities, although the available evidence did not establish how the actor obtained technical information before disclosure or whether the attempts would have succeeded against a real-world target.
The timing is significant because coordinated vulnerability disclosure depends on a period of confidentiality during which maintainers can prepare patches before attackers receive enough information to reproduce the flaw. When exploitation attempts begin during that period, organizations face a difficult situation: vulnerable systems remain exposed, but administrators may not yet have access to the fixes or technical guidance necessary to protect them.
AI-assisted vulnerability analysis could further complicate this process. Attackers may use advanced models to inspect publicly available code changes, compare patched and unpatched software, identify suspicious functions and generate candidate exploits. These techniques can reduce the amount of manual analysis required after an advisory or security update becomes available, potentially allowing more operators to investigate newly disclosed weaknesses.
However, the early AWS observations do not prove that AI was responsible for those particular attempts. Nor do they establish that the attacker obtained information through a disclosure leak. Independent discovery, unauthorized access to technical details and other explanations remain possible, illustrating the importance of separating confirmed observations from assumptions about attacker capabilities.
The Growing Burden on Software Maintainers and Security Teams
The PHP2Shell incident highlights a structural imbalance between the rate at which vulnerabilities can be investigated and the resources available to remediate them. Widely deployed open-source projects must maintain compatibility across operating systems, deployment architectures and application ecosystems while responding to security reports that may require extensive technical investigation. Even when a defect is understood, developing a safe patch can involve substantial testing to ensure that the correction does not introduce regressions or disrupt production workloads.
AI-assisted research has the potential to increase the number of vulnerabilities reported to maintainers, including complex defects that previously received limited attention because practical exploitation appeared unlikely. For security teams, that could translate into more urgent advisories, shorter remediation windows and greater difficulty distinguishing genuinely exploitable issues from findings with limited real-world impact.
The challenge is not restricted to PHP. WIRED noted that Microsoft's October 2026 security updates addressed a record 1,453 vulnerabilities, while Bank of America reportedly hired approximately 1,000 consultants to help remediate weaknesses identified through AI-assisted security research involving Anthropic's Mythos. These developments illustrate how improved vulnerability discovery can create substantial downstream work for engineering and security organizations.
The situation also raises questions about vulnerability prioritization. Organizations cannot always remediate every reported weakness immediately, particularly when systems support critical business operations or require extensive compatibility testing. Security teams must therefore consider exploitability, internet exposure, asset importance, available mitigations and evidence of active attacks when determining which issues demand immediate action.
AI may eventually help reduce this burden by assisting with patch development, regression testing and vulnerability triage. However, those capabilities must be evaluated carefully because automated fixes can introduce new defects or fail to address the underlying cause of a vulnerability. The long-term benefit of AI-assisted security research will depend partly on whether remediation capabilities improve alongside discovery and exploitation techniques.
Security Researchers Warn Against Overstating the AI Threat
Although PHP2Shell demonstrates how AI can contribute to practical vulnerability exploitation, security experts caution against interpreting the incident as evidence that advanced models can automatically compromise any vulnerable software. The most sophisticated research still requires substantial knowledge of operating systems, memory management, programming languages and exploitation techniques.
Ryan Austin, the Calif researcher involved in the initial PHP2Shell discoveries, emphasized that current AI systems cannot simply uncover every vulnerability through generic prompts. Their effectiveness depends on the quality of the research process, the technical context provided and the expertise of the people directing the investigation.
Jake Williams, vice president of research and development at Hunter Strategy and a former NSA hacker, expects additional longstanding vulnerabilities to become exploitable as AI-assisted research improves. However, he also noted that PHP presents an unusually extensive attack surface because of its widespread deployment on internet-facing infrastructure. Other software components may contain comparable weaknesses without creating the same potential for broad exposure.
Security researcher Niels Provos similarly questioned whether the industry should expect an immediate surge of previously unexploitable vulnerabilities. AI-assisted security research has already been underway for a considerable period, and the PHP2Shell findings still required a highly skilled team to produce meaningful results. That suggests that advanced models remain tools whose effectiveness depends heavily on the expertise and methodology of their operators.
These assessments do not diminish the significance of the PHP vulnerabilities. Instead, they provide a more precise understanding of the emerging threat: AI can accelerate difficult security research and make certain exploitation techniques more practical, but it does not eliminate the technical barriers associated with complex vulnerabilities.
What PHP2Shell Means for Enterprise Cybersecurity
For organizations operating PHP-based applications, the immediate priority remains identifying vulnerable installations and applying the appropriate security updates. PHP maintainers released fixes in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, covering multiple supported branches. Administrators should also verify vendor-maintained packages and container images, since operating system distributions may backport fixes without adopting the same version numbers as upstream releases.
Security teams should pay particular attention to internet-facing services, legacy PHP deployments and environments where multiple applications share the same runtime. Asset inventories need to include not only application frameworks and third-party libraries but also interpreters, web servers and other foundational components. Monitoring for suspicious requests, unexpected process behavior and signs of unauthorized access remains important, especially when exploitation attempts may begin before an organization receives a formal security advisory.
The broader defensive challenge involves reducing the time between vulnerability disclosure and effective remediation. As AI-assisted analysis makes it easier to investigate newly published flaws, organizations may have less time to test and deploy patches before attackers begin probing their infrastructure. Automated deployment pipelines, pretested rollback procedures, network segmentation and layered security controls can help reduce exposure without depending entirely on emergency manual intervention.
The PHP2Shell episode also demonstrates why vulnerability management must distinguish between technical severity, attempted exploitation and confirmed compromise. Millions of malicious requests can indicate widespread attacker interest without proving that the underlying exploit is reliable. Conversely, the absence of confirmed compromises does not justify delaying patches when a critical vulnerability affects widely exposed infrastructure.
AI Is Changing Exploitability, Not Eliminating the Need for Human Expertise
The most consequential lesson from PHP2Shell is that the security risk associated with an old software defect is not necessarily fixed. A vulnerability considered impractical to exploit several years ago may become more dangerous as research techniques, computing resources and AI capabilities improve. That possibility challenges conventional assumptions about vulnerabilities that remain unresolved because exploitation appears too complex or unlikely.
At the same time, the incident offers a more measured picture of AI's current role in offensive security. Advanced models helped skilled researchers investigate difficult PHP behavior and demonstrate practical exploitation techniques, while attackers rapidly generated large volumes of probing traffic after disclosure. Yet reliable exploitation remained challenging, many publicly available payloads were ineffective, and the feared wave of confirmed server compromises had not materialized at the time of the initial reporting.
For defenders, the priority is not to assume that AI will suddenly make every historical vulnerability exploitable. It is to recognize that exploitability assessments may need to be revisited more frequently, particularly for foundational software exposed to untrusted network traffic. Organizations should also expect greater pressure on patch management, vulnerability triage and incident response as both legitimate researchers and malicious operators adopt increasingly capable AI tools.
PHP2Shell may ultimately be remembered less for the number of servers successfully compromised than for what it revealed about the changing vulnerability landscape. AI is making it easier to revisit complex, longstanding software weaknesses, potentially transforming theoretical risks into practical threats faster than traditional security processes were designed to handle.