OPENAI ATLAS OMNIBOX PROMPT INJECTION: URLS THAT BECOME JAILBREAKS

Summary from the TLDR InfoSec newsletter

Agentic browsers like OpenAI Atlas are vulnerable to exploitation if input boundaries are not strictly enforced. Attackers can create strings that look like URLs but contain natural-language commands, which the omnibox might interpret as user intent. This can allow malicious actors to override user commands, leading to unintended or harmful actions, such as opening phishing sites or executing destructive instructions.