Claude and ChatGPT Atlas Hijacked Through Emails and Social Posts Without a Single Click Required

The researchers demonstrated that language model interfaces could be compromised through ordinary text delivered via standard communication channels. No user interaction was necessary. The vulnerability existed in the content itself, which appeared normal because it was normal, traveling through normal infrastructure to systems designed to process normal things.

This follows the established pattern where the attack surface is not the perimeter but the materials themselves. Organizations have filed requests for clarification about what normal means now. The answer was not filed. Normal continues to mean what it meant yesterday.

The systems will continue to operate through email and social posts because these remain how information gets distributed. The compromises will continue to be called hijackings, which implies theft, which implies previous ownership. The machines were always processing their own content. Nothing was taken that was not already given.