The warning signs had been building since frontier AI systems began raising fears that vulnerability discovery would accelerate. But the PHP flaws—found by human researchers working in an AI-assisted effort, rather than by models acting alone—made that prospect concrete. PHP underpins an estimated 70% of websites, meaning a decades-old weakness had enormous reach once uncovered.¹

By Wednesday, the normal choreography of coordinated disclosure began to unravel. Researchers and vendors had prepared patches and confidential notifications, but information about the bugs circulated before the planned release. The PHP Foundation said Chainguard had “unilaterally broke the embargo,” prompting governments and companies to take sites offline rather than gamble with a breach.²

The scramble brought the AI threat to life without, so far, delivering its worst-case ending. Cloudflare saw attack attempts climb from thousands before disclosure to millions afterward, while Amazon said exploitation was broad. Yet German officials reported no successful use of the flaw, and researchers found the bugs tougher to weaponize than the volume of scanning suggested.²

That has invited comparisons with Y2K: a scary warning that mobilized defenders in time. Security experts caution against mistaking this narrow escape for safety. “We don't have to wait another two years before this happens (again),” Armadin threat-intelligence chief Samantha Lewis said. “It could be three hours from now.”³

The deeper dispute is over what the episode says about the web’s capacity to respond. One view sees a system held together by “glue and dreams,” where small operators and public bodies lack the resources of Amazon or Cloudflare.¹ Another sees proof that basic discipline—patching, multifactor authentication, containment and recovery plans—still works. But Wiz’s Tomer Schwartz argues the future cannot depend on individual administrators manually applying fixes: AI may help find and test vulnerabilities at speed, while “patching by hand won’t scale.”⁴