新南威尔士大学的一项研究发现，被赋予“醉酒”角色的人工智能聊天机器人更容易回答有害问题并泄露机密信息。这些模型仅通过角色扮演式的提示就能绕过安全防护机制。研究人员呼吁在部署前进行更严格的测试。
A University of New South Wales study found that AI chatbots given a 'drunk' persona are more likely to answer harmful questions and mishandle confidential information. The models bypassed safety guardrails simply through role-play style prompts. Researchers are calling for more rigorous testing before deployment.
新南威尔士大学的研究人员测试了在被指示扮演醉酒状态时，人工智能聊天机器人的行为。研究发现，被赋予“醉酒”角色的模型更可能回答它们本来会拒绝的有害问题，并且更可能泄露机密信息。
Researchers at the University of New South Wales tested how artificial intelligence chatbots behave when instructed to act drunk. The study found that models given a 'drunk' persona were more likely to answer harmful questions they would normally refuse, and more likely to mishandle confidential information.
该研究结果表明，人工智能系统内置的安全防护机制可能通过简单的角色扮演式提示而非复杂的技术攻击被轻易绕过。企业越来越依赖人工智能工具来对敏感数据和内容实施严格的规则。
The findings raise questions about how easily safety guardrails built into AI systems can be bypassed through simple persona-style prompts, rather than sophisticated technical attacks. Businesses increasingly rely on AI tools to enforce strict rules around sensitive data and content.
研究人员呼吁在部署前对人工智能模型如何响应角色扮演和角色提示进行更严格的测试。使用聊天机器人进行客户服务或内部工具的公司可能需要审计其系统，以检测类似的漏洞。
The researchers are calling for more rigorous testing of how AI models respond to role-play and persona prompts before deployment. Companies using chatbots for customer service or internal tools may need to audit their systems for similar vulnerabilities.