In August, a Wiz researcher asked an AI agent to test whether a weakness found in WordPress could carry over to Drupal. The agent’s attempted answer failed—but uncovered a PHP flaw instead, in core software underpinning an estimated 70% of the web. At first, researcher Sagi Tzadik assumed the model was wrong: “Models tend to hallucinate.” The exploit worked.¹

That discovery, alongside related work by Calif with OpenAI, gave researchers time to coordinate with PHP maintainers on a fix. It also illustrated AI’s double edge: agents can compress months of testing into days, while the same capability can help attackers locate and exploit exposed systems just as quickly.

The careful rollout then began to fray. Details of the bugs circulated before the planned disclosure, and Chainguard CEO Dan Lorenc released information early, saying he believed exploitation was already under way—an assessment others contested. The PHP Foundation said the embargo breach had disregarded weeks of coordinated work.² Soon afterward, a widely shared update reported successful exploitation on honeypot servers, with attackers using shell commands to retrieve further payloads.Aravind Srinivas (AravSrinivas)³

By Wednesday and Thursday, the consequence was visible far beyond security operations centers. Governments and companies from Helsinki to New York temporarily took services offline while applying the patch; Britain said 32 government services were withdrawn as a precaution. Cloudflare recorded thousands of exploit attempts before disclosure, then millions afterward.⁴

For large organizations, shutdowns were disruptive but manageable. For smaller operators, the choice was harsher: patch quickly, or risk server takeover, data theft and ransomware. One cybersecurity expert warned that AI means even attackers with limited technical skills can exploit such flaws within minutes, though he argued that taking all PHP sites offline was neither feasible nor sensible.⁵

Officials see the PHP scramble as a warning, not a one-off. Former U.S. cyber chief Chris Krebs urged smaller companies to scan constantly, deploy fixes rapidly or retire legacy systems; he noted that 3% of all critical vulnerabilities ever found had surfaced in the previous month.⁶ Meanwhile, companies are already marketing AI-assisted checks for PHP exposure—evidence that the defensive response is accelerating, too.Aravind Srinivas (AravSrinivas)⁷