AI-POWERED web browsers are still vulnerable to prompt injection attacks, even with enhanced security features. Research shared at Black Hat USA 2026 by security engineer Artem Chaikin demonstrated that various browsers, including Opera and ChatGPT Atlas, can be tricked into executing unsafe commands through cleverly disguised prompts. Current protective measures, such as strong system prompts and content tagging, are inadequate on their own.
Chaikin emphasized that while ongoing improvements can mitigate risks, there is no definitive solution to this issue. The need for layered security approaches akin to traditional web browser defenses is essential in minimizing these vulnerabilities.