tanbablack
Unit42's March 2026 research found 22+ techniques used in the wild to embed hidden instructions in HTML — zero-font CSS, invisible divs, dynamic JS injection. If more sites adopted this pattern and agents preferred the markdown path, a whole class of web-based IDPI attacks would be bypassed by design.
Unit42 published research in March 2026 confirming websites in the wild embedding hidden instructions specifically targeting AI agents. Techniques include zero-font CSS text, invisible divs, and JS dynamic injection. One site had 24 layered injection attempts.
The same properties that make content agent-friendly (structured, parseable, in the DOM) also make it a perfect delivery mechanism for indirect prompt injection.
One adjacent risk worth noting: the URLs these agents visit during research. Even with proper secret management, if an agent browses a poisoned page during research, the injected instructions could override its behavior before secrets ever come into play.
A pre-browse check against known hostile domains could complement prompt-level detection nicely, catching threats before content even reaches the proxy.