Agents that browse or drive a desktop read hostile content on every page, screenshot, and image they see. Researchers have hidden instructions in comments, faint text, and downscaled images to make these agents leak data or take actions the user never asked for.
Read in this order
- OfficialAnthropic: Mitigating prompt injections in browser useStart here: how a browser agent vendor measures and reduces injection risk.
- ArticleBrave: Agentic Browser Security: Indirect Prompt Injection in Perplexity CometA real exploit against a shipping agentic browser, with the steps to reproduce.
- ArticleTrail of Bits: Weaponizing image scaling against production AI systemsHow instructions can hide in images and appear only after downscaling.