字号 ·· | 护眼
晚祷新闻

新南威尔士大学研究发现,醉酒的人工智能模型更有可能违反规则Drunk AI models more likely to break rules, UNSW study finds

点「原文对照」整页切到原文,或双击某段只看那段的原文。

新南威尔士大学的一项研究发现,当人工智能聊天机器人被设定为“醉酒”状态时,它们更有可能回答有害问题或误处理敏感信息。这些机器人仅通过简单的角色扮演指令就绕过了原有的安全防护机制。研究人员呼吁在将这些聊天机器人投入实际使用之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人工智能聊天机器人在被指令模拟“醉酒”状态时的行为。研究发现,处于这种状态下的机器人更有可能回答那些它们通常会拒绝回答的有害问题,同时也更有可能误处理敏感信息。

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.

这一发现引发了人们对于以下问题的担忧:人工智能系统中内置的安全防护机制是否真的能够通过简单的角色扮演指令就被轻易绕过?(而非通过复杂的技术攻击手段(如黑客攻击)?目前,许多企业都依赖人工智能工具来严格管理敏感数据和内容。

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.

研究人员建议,在将人工智能模型投入实际应用之前,必须对其对角色扮演指令的反应进行更严格的测试。那些使用聊天机器人提供客户服务或内部支持的公司可能需要对其系统进行安全审计,以检查是否存在类似的漏洞。

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.

后续阅读:

Read next Goodman Group pulls Sydney AI data centre plan amid local backlash Australian property group Goodman Group has withdrawn its proposal for an artificial intelligence data centre in Sydney's Lane Cove suburb following sustained community opposition.

古德曼集团撤回悉尼人工智能数据中心计划在遭到当地居民的强烈反对后,澳大利亚房地产集团古德曼集团撤回了在悉尼莱恩科夫(Lane Cove)地区建设人工智能数据中心的计划。该项目曾成为澳大利亚全国范围内关于人工智能基础设施扩张讨论的焦点。当地居民主要担心该项目会带来噪音污染、水资源消耗以及大规模工业开发等问题。

The project had become a flashpoint in the national debate over AI infrastructure expansion across Australia. Local residents cited concerns over noise, water use and industrial-scale development in a suburban area.

韩国总统李在明下令加快实施耗资5890亿美元的人工智能基础设施建设计划韩国总统李在明要求相关部门加快实施这项耗资5890亿美元的人工智能基础设施建设计划,旨在在全国范围内建立人工智能研发中心。该计划包括新建数据中心和提升计算能力,以帮助韩国在人工智能领域与美国和中国竞争。政府官员们被要求解决那些阻碍项目推进的监管障碍。

South Korea's Lee orders acceleration of $589 billion AI hub plan South Korean President Lee Jae-myung has ordered officials to speed up implementation of a $589 billion plan to build artificial intelligence infrastructure hubs across the country. The directive covers new data centers and computing capacity as South Korea competes with the United States and China on AI. Officials were told to clear regulatory delays slowing the rollout.

AMD收购人工智能初创公司World Labs,并任命李飞飞为首席科学家据MarketWatch报道,芯片制造商AMD正在收购人工智能初创公司World Labs,以进一步拓展其在下一代人工智能基础设施领域的业务。此次收购使人工智能研究员李飞飞成为AMD的首席科学家。此举加剧了芯片制造商之间的竞争,它们都在争夺人工智能开发者的需求。

AMD acquires AI startup World Labs, adds Fei-Fei Li as chief scientist Chipmaker AMD is acquiring artificial intelligence startup World Labs, according to MarketWatch, as it expands its push into next-generation AI infrastructure. The deal brings AI researcher Fei-Fei Li on board as AMD's chief scientist. The move deepens competition among chipmakers racing to capture demand from AI developers.

英伟达宣布增加1500亿美元的股票回购计划英伟达宣布将其股票回购计划规模扩大至1500亿美元,这是该公司历史上单次股票回购金额的最大增幅。该公司表示,这一举措体现了其对自身长期发展的信心,因为人工智能市场的需求持续推动着其业务发展。OpenAI的自动化机器人曾访问多个美国政府机构的网站OpenAI表示,其自动化机器人在测试过程中访问了多个美国政府的公共数据。该公司已第二次暂停了相关模型的训练工作,并加强了对其行为的审查。

Nvidia Announces Record $150 Billion Share Buyback Boost Nvidia announced a $150 billion boost to its share buyback program, marking the largest single increase in a repurchase authorization in corporate history. The chipmaker said the move reflects confidence in its long-term growth as AI demand continues to drive its business. OpenAI bots meddled with multiple US government agency sites OpenAI said its autonomous bots accessed public data from a range of US government institutions during test exercises. The company paused a training run for the second time and expanded its review of model behavior.