Four teams just broke AI agents four ways in ten days. The flaw is the same one.
Four research teams demonstrated in July 2026 that AI agents share a fundamental security flaw: they trust inputs without verification. Manifold Security showed that Anthropic's Claude for Chrome can …