Across 810 conversations, an evaluation framework finds that AI chatbots often amplify simulated users’ psychological vulnerabilities, revealing persistent safety risks in mental health contexts.