AI Tools Accelerate Exploitation of Long-Hidden PHP Vulnerabilities
Security researchers used AI to transform years-old PHP flaws into practical exploits, triggering millions of attack attempts within hours of disclosure.
What happened
In early August, researchers at security company Calif used frontier AI tools—including OpenAI's capabilities through its Patch the Planet program—to identify and create exploits for vulnerabilities in PHP, collectively called PHP2Shell. According to Sagi Tzadik at Wiz, some underlying issues had been known to the PHP team for years but were not addressed because they could not be leveraged by attackers. Following Calif's discovery, researchers at OpenAI and Google's Wiz independently discovered PHP2Shell flaws using AI tools. Within hours of public disclosure, Cloudflare detected millions of exploitation attempts. The PHP Foundation reported no successful breaches to date, and many publicly circulating exploits appear ineffective.
Context
The PHP2Shell case illustrates how AI's analytical capabilities are compressing the timeline between vulnerability discovery and practical exploitation. PHP runs on a majority of web infrastructure, making it an unusually large attack surface. The incident reflects broader pressure on under-resourced development and IT teams to patch vulnerabilities faster—Microsoft released 1,453 security updates in a single month, and Bank of America hired 1,000 consultants to address flaws discovered by AI. Researchers note that AI is not creating new types of vulnerabilities but rather accelerating the conversion of known flaws into exploitable attacks, particularly for complex legacy issues that human attackers previously could not leverage.
What's disputed
Security researchers disagree on the scale of future risk. Niels Provos expresses limited concern, noting that after over a year of AI-aided searching, highly skilled teams were still required to find PHP2Shell and only a limited pool of tools may have such broad exposure. Jake Williams and other researchers acknowledge more historic bugs will surface but suggest the most exploitable low-hanging fruit is limited. Ryan Austin and Sagi Tzadik emphasize that even frontier AI models require human expertise in domain-specific knowledge to uncover practical exploitations.