How AI Accelerated a Long-Hidden PHP Security Crisis

A cluster of critical vulnerabilities in PHP—collectively called PHP2Shell—demonstrated a troubling new dynamic in digital security: artificial intelligence capable of transforming dormant flaws into active threats. Within hours of disclosure, hackers mounted millions of exploitation attempts, yet the role of AI in making these historic bugs dangerous offers a more nuanced picture than panic alone suggests.

What Happened#

Security researchers at Calif discovered the PHP2Shell vulnerabilities in August using frontier AI models, including OpenAI's tools as part of its Patch the Planet initiative. The flaws had sat in PHP's codebase for years—some dating back to 2024—but were dismissed as impractical to exploit. Following Calif's work, researchers at OpenAI and the security firm Wiz independently leveraged AI to rediscover the same issues. Though millions of automated attacks emerged after the patches went public, the PHP Foundation reported no confirmed successful breaches using these flaws, and many circulating exploits appear to be ineffective.

AI's Role in Acceleration#

According to Sagi Tzadik, a Wiz researcher, "The vulnerability had been known to the PHP team for years, but because it could not be leveraged by an attacker—at least, not a human one—it was not addressed." Advanced AI models compressed the time needed to convert theoretical bugs into working attacks, amplifying pressure on already-stretched security teams. The effect arrived at a particularly acute moment: Microsoft alone reported 1,453 security vulnerabilities in October, while Bank of America recently contracted 1,000 consultants to remediate flaws Anthropic's systems had surfaced.

What Experts Say#

Not all researchers share doomsday expectations. Jake Williams, vice president of research at Hunter Strategy and a former NSA hacker, notes that while long-buried vulnerabilities may surface periodically, the opportunity set is limited. "There isn't much else that has the specific reach of PHP," he observes. "Most other tools aren't exposed to the internet by design." Niels Provos, a security researcher, adds measured skepticism: despite a year of AI-aided searching, finding this one vulnerability required a highly skilled team and sophisticated expertise—a pattern suggesting the threat, while real, remains constrained.

Critically, even breakthrough AI hacking still demands human knowledge. As Ryan Austin, the Calif researcher who first unearthed PHP2Shell, wrote: "No model is simply capable of unearthing 'all bugs' with simple prompts. At least for now."