AI-assisted researchers have uncovered two critical PHP vulnerabilities that could put as many as 70% of websites at risk, prompting an emergency patch and warnings that attackers may exploit the flaws before organisations can update their systems.
The vulnerabilities, identified as CVE-2026-109618 and CVE-2026-109619 and collectively dubbed “php2shell”, allow attackers to execute code remotely on affected PHP systems. That could provide an entry point into the wider networks hosting those machines, researchers said.
The PHP Foundation, which maintains the open-source programming language, urged website operators to apply the patch immediately or take their services offline until they can do so.
Cloudflare, which helps protect millions of websites, said it had already observed “early attacks” as reports of the flaws emerged before the official release of the fix. Successful breaches remain limited and are still being assessed.
One startup owner alleged that attackers had stolen source code, exposed cookie data and leaked legal documents after exploiting the PHP bug, describing the vulnerability as “real bad” and saying people should be “afraid”.
Several governments have taken precautionary action by shutting down websites, including agencies in the Netherlands and Finland. The FBI’s Cyber Division, the NSA, the Japanese prime minister’s office and the European Union’s Computer Emergency Response Team have all urged potentially affected organisations to update their systems.
Gang Wang, an associate professor of computer science at the University of Illinois’ Grainger Engineering, said the vulnerabilities could expose a large number of websites to immediate attacks if they were not patched in time.
David Nicol, director of the university’s Information Trust Institute, said attackers could use the flaws as an entry point into the network hosting a vulnerable machine and potentially carry out attacks causing significant damage.
PHP is used across much of the web. WordPress, which runs on the language, powers roughly 40% of all websites, while other sites identified as using PHP include Microsoft.com, Wikipedia.org, Archive.org, Theguardian.com and Baidu.com.
An initial assessment by Palo Alto Networks identified 22.6 million potentially vulnerable instances. Facebook also began as a PHP website, although the company started a major move to different software in 2020.
How AI tools found the PHP vulnerabilities
The flaws were independently identified by several security organisations in August, with AI tools used in each investigation. The groups then worked with the PHP Foundation to develop a fix and coordinate its release.
The first discovery came from Calif, a California-based AI red-teaming firm involved in OpenAI’s Patch the Planet initiative, which seeks to find and repair vulnerabilities in digital infrastructure. The team began examining PHP because it considered the language a high-impact target.
Using OpenAI’s GPT-5.6 Sol model, Calif found the first issue on 1 August and alerted OpenAI. Further investigation by OpenAI uncovered additional elements, revealing a more serious and comprehensive vulnerability than initially understood.
The underlying bug had been present in PHP for 14 years. A limited version had been discovered in 2024 but was not considered significant at the time, according to OpenAI co-founder and president Greg Brockman.
Wiz, a Google-owned cybersecurity company, independently identified the first issue on 19 August using Google’s Gemini 4 Argon, OpenAI’s GPT-5.6 Sol and Wiz Atlas. It had focused on PHP because of its central role in internet infrastructure.
OpenAI subsequently brought Calif, Wiz, the PHP Foundation, Cloudflare, Quarkslab and several US government agencies into a co-ordinated effort to address the vulnerabilities and manage the disclosure.
OpenAI said the incident demonstrated how AI could uncover weaknesses with internet-scale consequences, while warning that the same capabilities could help attackers develop offensive tools and exploits more rapidly.
