Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents - Scale Labs | Scale Labs