PHP2Shell: AI opens 10-year-old PHP flaw used by 70% of web

The PHP2Shell incident did not stem from a recent coding mistake. According to the PHP2Shell official site and Nomura Research Institute, the flaw had sat deep in PHP for more than a decade because exploiting it required combining conditions so complex that human hands and heads were thought practically incapable of doing so. Wiz Research reported that one researcher, working with cutting-edge AI, cracked that puzzle in a week—a task an elite human team might spend months on. The episode shows that a safety margin long taken for granted, one resting on human limits rather than deliberate design, has quietly disappeared.

Anthropic had foreshadowed this shift. In April 2026 it developed Claude Mythos, found it too capable for open release, and restricted it to defense experts, warning that similar abilities would eventually reach those who do not intend to use them safely. By late September 2026, Anthropic Research reported that the publicly available GLM-5.3 came close to Claude Mythos Preview's performance in experiments to create attack programs. Meanwhile, PHP2Shell itself was discovered and demonstrated by defensive researchers using Gemini and GPT-family AI—a reminder that the same capabilities serve both sides.

The technology is separate from its abuse. Distillation lets a new model learn efficiently from a stronger teacher's outputs, and open-weight models publish their inner parameters for anyone to download and modify. These practices carry real value for research and industry, but once a model is open-weight, its built-in guardrails can be removed. Anthropic warns this makes it easier for attackers to misuse advanced techniques. The ethical thread runs back to the fact that humans themselves built AI surpassing human ability, opening a Pandora's box that cannot be closed. The practical response it suggests: inventory the systems you depend on, separate critical back-end functions from public-facing interfaces—as Japanese banks did when they temporarily shut PHP-based web pages but kept payment cores running on separate systems—and have humans ready to decide when to stop a service before damage spreads.