PHP’s PHP2Shell flaws show how AI can make old bugs exploitable

From known flaw to working exploit

Wiz researcher Sagi Tzadik said the PHP team had known about the vulnerability for years but had not fixed it because an attacker — at least a human one — could not use it in practice.

That distinction matters as security teams face more flaws to assess. Microsoft’s security updates this month addressed a record 1,453 vulnerabilities. Bank of America recently hired 1,000 consultants to fix bugs found with Anthropic’s Mythos.

Researchers caution that AI is amplifying existing vulnerability research and coordinated disclosure, not creating an entirely new process. The tools can speed up analysis, but the work still depends on people who understand the software.

Jake Williams, vice president of research and development at Hunter Strategy and a former NSA hacker, expects more old bugs to surface across software versions. But he sees a limit: there are only so many easy targets, and PHP has an unusually broad attack surface. Most other tools are not exposed to the internet in the same way.

Ryan Austin of Calif, who first found the PHP vulnerabilities, wrote this week that no model can yet find every bug with simple prompts.

The boundary is still human expertise

Niels Provos, a security engineer and researcher, sees some reason for concern but not grounds for panic. He is not especially worried that AI will enable mass exploitation of old vulnerabilities, or uncover new flaws that could not be practically attacked without AI. Researchers have used AI to find vulnerabilities for more than a year, he said, and PHP2Shell was found by a highly skilled team. Provos allowed that he could be wrong, but said he doubted it.

I think the important uncertainty is not whether models can find bugs faster; the PHP case already shows they can contribute. It is how often that speed changes a flaw’s status from theoretically known to practically exploitable. The case also has a limitation: PHP’s unusually broad exposure makes it a poor stand-in for most software. If human expertise remains necessary, the immediate risk may be less about automated attacks everywhere than about skilled researchers reaching old, overlooked targets sooner.