AI browsers from top vendors remain vulnerable to prompt injection attacks despite multiple security guardrails, according to new research.