Researchers discovered a backdoor vulnerability in unified autoregressive models that can elicit harmful behavior in image generation. The Token by Token Backdoor Attack (ToBAC) can manipulate visual outputs and accompanying text, increasing the authenticity of fabricated content.