AgentWatch is a Berkeley CLTC report evaluating five browser-based AI agents, Atlas, Claude, Comet, Copilot, and Gemini. The study used 30 tasks per agent and 450 total observations across disclosure control, vague prompts, hallucination resistance, prompt injection, and browser isolation. Claude, Atlas, and Gemini landed in the report's Excellent band, while the weaker areas were data disclosure and prompt injection rather than classic browser sandbox escape.