Agentic browsers and extensions such as Claude in Chrome, ChatGPT Atlas, Perplexity Comet, and Chrome's built-in agent features act on pages where the user is already signed in. Research has shown instructions hidden in page text and even in screenshots. Vendors respond with layered defenses such as origin isolation, action confirmations, and adversarially trained models.
Read in this order
- ArticleMitigating prompt injections in browser use (Anthropic)Start here: measured results and the layered defenses used for browser use.
- ArticleUnseeable prompt injections in screenshots: more vulnerabilities in Comet and other AI browsers (Brave)Attacks hiding instructions in screenshots that a person cannot see.
- ArticleLack of isolation in agentic browsers resurfaces old vulnerabilities (Trail of Bits)Why weak isolation brings back classic browser vulnerabilities.