Prompt Injection and Data Theft Risks Emerge In Agentic LLM Browsers

The report warns that agentic LLM browsers (e.g., Perplexity Comet, OpenAI Atlas, Edge Copilot, Brave Leo) expose new attack surfaces where prompt injection and XSS can cross privileged browser-LLM bridges to hijack AI agents, perform actions across tabs and local resources, and silently exfiltrate sensitive data; although some mitigations exist, the fundamental need for agentic browsers to cross long‑hardened boundaries makes these high‑impact risks difficult to fully eliminate.