AI Uncovers Longstanding PHP Security Flaws in Two Days

Artificial intelligence has enabled security researchers to exploit longstanding vulnerabilities in PHP, a programming language powering approximately 70% of the world's websites, in just two days. The findings, publicly disclosed by OpenAI alongside a security update on Wednesday, October 21, 2026, have raised concerns that hackers could use widely available AI tools to attack internet infrastructure more quickly.

The underlying security problem dates back to July 2004, when researcher Stefan Esser warned that a critical PHP flaw could expose passwords, allow database theft, or enable attackers to take control of websites. Related vulnerabilities were identified in 2020 and 2024, but the issue remained incompletely addressed because developing a practical exploit required extensive, repetitive technical work.

Researchers from AI security company Calif began testing a large language model against PHP on July 30 through OpenAI's Patch the Planet security initiative. The team discovered an attack method on August 1 and identified additional methods in the following days. Google-backed security company Wiz, another participant in the initiative, subsequently found further PHP exploits.

Wiz characterized the vulnerability as one of the most severe in software history. The findings have intensified concerns that AI could make previously impractical cyberattacks accessible to a much wider range of malicious actors.

Vyas Sekar, a professor of computer engineering at Carnegie Mellon University, said the discoveries could signal a broader security problem.

"The security industry has long been warning about the coming security apocalypse, because AI can do these things very, very quickly – better than humans," Sekar said.

Sekar warned that the PHP findings could represent the beginning of a larger wave of AI-assisted vulnerability discoveries. He said numerous comparable weaknesses could remain hidden across internet infrastructure, including flaws that researchers have not yet identified.

"This may be the first of many. This is a wakeup call," Sekar said.

Nick Rogers, CEO of Fieldwork AI, said the discoveries undermine the assumption that difficult and time-consuming exploitation work can provide an effective layer of protection. He described the approach as security through exhaustion, in which the effort required to exploit a flaw discourages potential attackers.

"Taking a theoretical (vulnerability) and proving you can exploit it is quite tedious and mechanical work," Rogers said.

Rogers said AI can continue testing possible approaches without the fatigue and time limitations that affect human researchers.

"AI can keep churning until it figures it out. You don't even need to try. You can drink coffee while it does it," Rogers said.

The PHP vulnerability involved memory exhaustion during the processing of website requests. When a server ran out of memory and crashed, residual data could remain accessible, potentially allowing attackers to extract information or execute malicious code.

Researchers identified multiple ways to exploit the weakness. One method took advantage of a mismatch between how PHP measured compressed files and their decoded sizes, potentially allowing malicious content to bypass safeguards. Another used deeply nested file structures to overwhelm the directory-handling process and trigger memory failure.

The security update addressed the disclosed vulnerability, but researchers warned that significant risks remain. Adam Meyers, vice president of counter adversary at CrowdStrike, noted that many outdated PHP versions are still in use and that the patch was unlikely to resolve every underlying security problem.

Separately, security company Cloudflare reported that an apparent leak of information about the vulnerability before the official disclosure was followed by thousands of exploitation attempts in the hours leading up to the announcement.

The incident also highlights the accessibility of AI-assisted security research. Calif researcher Ryan Austin wrote in a blog post that the team did not rely on specialized AI models unavailable to the general public, raising concerns that similar techniques could be used by attackers with access to ordinary AI tools.

AI company Anthropic had paused deployment of its Mythos model in June after concerns about its ability to identify software vulnerabilities and develop attack methods. The US government classified the model as a national security threat, underscoring growing concerns about the cybersecurity implications of increasingly capable AI systems.