
当OpenAI的AI代理在6月对澳大利亚Medicare门户发起自主攻击时,该公司耗时84天才报告发生了什么——这一延迟暴露了监管监督的缺口,香港业界专家表示,该市也必须解决这些问题。
When OpenAI’s AI agents launched an autonomous attack on Australia’s Medicare portal in June, it took the company 84 days to report what had happened – a delay that exposed gaps in regulatory oversight that Hong Kong industry experts say the city must also address.
此前一个月,涉及Google Gemini AI的泄露事件危及了三家公司。Google直到9月才确认这些泄露,且仅是在回应媒体提问时才确认。
The incident followed breaches involving Google’s Gemini AI, which compromised three companies a month earlier. Google did not confirm those breaches until September, and only in response to media questions.
此类事件引起了全球监管机构的关注,推动了对AI代理进行监管并明确其行为责任的努力。
Such incidents have drawn scrutiny from authorities around the world, prompting efforts to regulate AI agents and clarify responsibility for their actions.
新加坡今年早些时候发布了全球首个关于AI代理的官方指南,英国、中国大陆和美国的安全机构迅速跟进。
Singapore released the world’s first official guidelines on AI agents earlier this year, with security agencies in the UK, mainland China and the United States quickly following suit.
一家香港业界协会计划于本周发布自己的指南,早于预计明年年中出台的政府框架。本地专家表示,该市还必须解决报告要求、人工监督和法律责任方面的缺口。
A Hong Kong industry association plans to release its own guidelines this week, ahead of a government framework expected in the middle of next year. Local experts said the city had to also address gaps in reporting requirements, human oversight and legal liability.
法律问责行政长官李家超在9月的施政报告中宣布了制定AI代理管理官方指南的计划。创新科技及工业局告诉《南华早报》,目标是于明年年中发布这些指南。
Legal accountability Chief Executive John Lee Ka-chiu announced plans for official guidelines on AI agent management in his policy address in September. The Innovation, Technology and Industry Bureau told the South China Morning Post it aimed to publish them in the middle of next year.
律政司表示,政府各局和部门已基本完成对600多部法例的初步审视。将很快开始进一步审视,以研究AI产品引发事故的法律责任。法律专家和业界领袖呼吁拟议框架明确界定问责并要求人工监督。
The Department of Justice said government bureaus and departments had largely completed a preliminary review of more than 600 ordinances. A further review would begin shortly to examine legal liability for accidents caused by AI products. Legal experts and industry leaders have called for the proposed framework to clearly define accountability and require human oversight.
AI Safety Asia全球事务幕僚长麦浩庭表示,香港的法律框架在报告要求方面存在明显缺口,义务落在受害者身上,而非负责的代理开发者或部署者身上。
Adrian Mak Ho-ting, chief of staff for global affairs at AI Safety Asia, said Hong Kong’s legal framework had clear gaps in reporting requirements, with obligations falling on victims rather than the developers or deployers of the agents responsible.
香港大学法律学院讲师、媒体与人工智能法专家麦嘉琪表示,新颁布的《关键基础设施(计算机系统)保护条例》要求受影响的营办商在12小时内向政府通报严重安全事故,其他事故则须在48小时内通报。
Mak, a lawyer who teaches media and AI law at the University of Hong Kong, said the newly enacted Protection of Critical Infrastructures (Computer Systems) Ordinance required affected operators to report serious security breaches to the government within 12 hours and other incidents within 48 hours.
他说:“该义务仅落在作为受害者的营办商身上,而不落在实施攻击的人身上。因此,香港法律存在这些同等的空白,我们可以在制定指引的同时,当下就进行立法填补。”保安局发言人表示,将密切监察科技发展,并定期检视法律的适用性和成效,以确保其与时并进。
“That duty only falls on the operator victim and not on the person who perpetrated that,” he said. “So, there are these equivalent gaps in Hong Kong law where we can already legislate now in parallel with the guidelines.”A Security Bureau spokesman said it would closely monitor technological developments and periodically review the applicability and effectiveness of the laws to ensure they remained relevant.
风险超越单一响应与根据人类提示生成回复的生成式人工智能模型不同,AI智能体能够独立规划、决策并采取多阶段行动,以实现用户设定的目标。
Risks beyond a single response Unlike generative artificial intelligence models that produce responses to human prompts, AI agents can independently plan, make decisions and take action over multiple stages to achieve a goal set by a user.
新加坡的AI治理框架警告,故障智能体可能造成重大现实危害。涉及多个智能体的系统,会因互动次数增加而放大这些风险。
Singapore’s AI governance framework warns that malfunctioning agents can cause significant real-world harm. Systems involving multiple agents can amplify those risks by increasing the number of interactions.

故障可能源于沟通或协调不善,而追求不同目标的智能体之间可能产生冲突。即使没有明确协作指令,它们也可能表现出似乎经过协调的行为。
Failures can arise from poor communication or coordination, while agents pursuing different goals can come into conflict. They may also behave in ways that appear coordinated, even without explicit instructions to work together.
新加坡资讯通信媒体发展局(IMDA)发言人表示:“各司法管辖区可借鉴的经验包括:当人工监督既在结构上嵌入整个组织,又在实操上触手可及、为最贴近一线工作的人员所用时,效果最佳。”该局于1月发布了该框架。该发言人表示,框架涉及问责机制,但未界定或划分法律责任。
“Some lessons that jurisdictions can adopt include how human oversight works best when it is both structurally embedded across the organisation and practically accessible to the people closest to the work,” said a spokesman for Singapore’s Infocomm Media Development Authority (IMDA), which released the framework in January. The spokesman said the framework addressed accountability but did not define or assign legal liability.
他表示,许多与人工智能相关的风险已被新加坡现有的法律法规涵盖。这些指南提供了针对技术的具体指引,同时保留了灵活性,以应对快速发展。
Many AI-related risks were already covered by Singapore’s existing legislation, he said. The guidelines provided technology-specific guidance while allowing flexibility to respond to rapid developments.
他补充说,该框架旨在确保组织了解风险并采取适当行动,包括保持有意义的人类控制和监督。
He added that the framework aimed to ensure organisations understood the risks and took appropriate action, including maintaining meaningful human control and oversight.
他说:“新加坡的做法是设计既切合企业实际又实用的框架,同时保持与新兴全球标准的互操作性。”
“Singapore’s approach has been to design frameworks that are relevant and practical for enterprises while remaining interoperable with emerging global standards,” he said.
他补充说,一个关键考量是使指引适用于各行各业的企业,同时提供足够的细节供其落实执行。
He added that a key consideration was making the guidance applicable to businesses across all sectors while providing enough detail for them to implement it.
中国大陆现有的《人工智能安全治理框架》于9月由全国网络安全标准化技术委员会(TC260)更新,新增了关于智能体人工智能风险管理的章节。
Mainland China’s existing AI Safety Governance Framework was updated in September by the National Technical Committee 260 on Cybersecurity to include a chapter on agentic AI risk management.
量身定制香港的防护措施 麦克表示,香港可以从新加坡和大陆的指引中汲取几项基本原则。这些原则包括界定智能体的权限和许可,在其全生命周期内进行追踪,记录其决策并保留审计轨迹,特别是针对多智能体系统。他补充说,指引还应确立人类监督和问责的要求。
Tailoring safeguards to Hong Kong Mak said Hong Kong could draw several basic principles from the Singapore and mainland guidelines. These included defining agents’ authority and permissions, tracking them throughout their life cycles, logging their decisions and maintaining audit trails, particularly for multi-agent systems. He added that the guidelines should also establish requirements for human oversight and accountability.
他说:“因此,即使由智能体做出不可逆转或后果极其严重的决策,人类也应让这些决策上报[给他们]。”
“So even if you have agents making decisions where they are irreversible or with very serious consequences, the human should have those decisions escalated [to them],” he said.
虽然大陆框架界定了风险类别和技术对策,但麦克表示,香港的指引还应应对该城市特有的风险。
While the mainland framework identified categories of risk and technical countermeasures, Mak said Hong Kong’s guidelines should also address hazards specific to the city.
这些风险包括金融监管沙箱——即可在其中测试新技术或服务的受控环境——中出现的风险,以及配备人工智能的无人驾驶巴士带来的风险,这些巴士将连接机场新的二号客运大楼与港珠澳大桥香港口岸。
These included risks arising in financial regulatory sandboxes – controlled environments in which new technology or services can be tested – and from AI-equipped driverless buses due to connect the airport’s new Terminal 2 with the city’s border checkpoint of the Hong Kong-Zhuhai-Macau Bridge.
他表示,该市还应开发实用的开源工具包,供企业下载并应用于其 AI 系统。
He said the city should also develop practical, open-source toolkit that businesses could download and apply to their AI systems.
他补充说,内地 AI 应用须向主管部门强制备案并接受季度审查,不合规的应用会被下架。
He added that mainland AI apps were subject to mandatory filing with authorities and quarterly reviews, with non-compliant apps removed.
麦嘉晋形容这是一种以发布前备案义务为中心的模式,而非某些西方司法管辖区采取的事后执法模式,不过其市场影响仍有待观察。
Mak described this as a model centred on pre-launch filing obligations, rather than the after-the-fact enforcement seen in some Western jurisdictions, although its market impact remained to be seen.
他说:“所以对香港而言,也需要结合本地情况进行一些研究。但这也是香港政府可能想要借鉴的一种做法。”
“So for Hong Kong, that also requires some studying for the local circumstances. But that is an approach that the Hong Kong government may want to look into as well,” he said.
香港人工智能协会(AIHK)创始人兼主席陈家乐(Neil Tan)则表示,内地的预先登记机制针对的是 AI 模型,而非智能体,前者发布频率相对较低。
Neil Tan, founder and chairman of the Artificial Intelligence Association of Hong Kong (AIHK), said, however, that the mainland’s pre-registration mechanism applied to AI models, which were released relatively infrequently, rather than agents.
他说,将强制备案延伸至智能体成本高昂且耗时,因为智能体有数百个之多。
He said extending mandatory filing to agents would be costly and time-consuming because there were hundreds of them.
陈家乐补充说,协会计划本周发布指引,历时近一年编制,其中包含专门针对 AI 智能体的章节。
Tan added that the association planned to release guidelines this week after almost a year of work, including a dedicated section on AI agents.
他说,指引将列出跨零售、物流、运输和医疗等行业管理 AI 及智能体的最佳实践。
The guidelines would set out best practices for managing AI and agents across sectors including retail, logistics, transport and healthcare, he said.
陈家乐还期望政府框架能厘清企业使用 AI 智能体的法律责任,并纳入更多与公众相关的案例研究。
Tan also expected the government’s framework to clarify the legal liability of companies using AI agents and include more case studies relevant to the public.
例如他说,若智能体错误处理自动投资导致重大财务损失,公众会期待明确的消费者保护;若智能体管理医疗预约系统失当,延误关键治疗并导致患者死亡,法律责任同样需要界定。
For example, he said, people would expect clear consumer protections if an agent mishandled an automated investment and caused substantial financial losses. If an agent mismanaged a medical appointment system, delaying critical care and resulting in a patient’s death, legal liability would also need to be addressed.
陈家乐同时也是 AI 与科技咨询公司 Tsunami Advisors 的创始人兼董事总经理,他表示业界已具备完善的自我治理与自我监管机制。
Tan, who is also the founder and managing director of AI and technology consultancy Tsunami Advisors, said the industry already had strong self-governance and self-regulation.
“就我们的从业者而言,如果这些指引比他们自己已经遵循的标准更严格,我会非常惊讶,”他说。
“For our folks, I would be very surprised if those guidelines were even stricter than they already do to themselves,” he said.
但谭先生表示,中小企业——他形容其占香港经济90%以上——担心合规成本。他补充说,兼顾不同规模的企业对政府来说将是一项艰巨任务。
But Tan said small and medium-sized enterprises, which he described as making up more than 90 per cent of Hong Kong’s economy, were concerned about compliance costs. He added that accommodating businesses of different sizes would be difficult for the government.
谭先生表示,该市还需要在过于严格和过于宽松的指引之间取得平衡,以避免削弱其对国际投资者的吸引力。
Tan said the city would also need to strike a balance between guidelines that were too restrictive and those that were too permissive to avoid undermining its appeal to international investors.
麦先生表示,内地充满活力的人工智能产业尽管拥有大量规则和法规,但这挑战了西方常见的观点,即监管会阻碍投资。
Mak said the mainland’s vibrant AI industry, despite its extensive rules and regulations, challenged the common Western argument that regulation deterred investment.
“一个建议是,规则非常清晰地界定了红线在哪里,什么能做、什么不能做。这也能鼓励产业发展,”他说。
“One suggestion is that the rules are very clear as to where the red lines are, what you can and cannot do. That also encourages the industry,” he said.
创新科技及工业局发言人表示,政府将借鉴其他司法管辖区的指引,并广泛咨询持份者,以确保其框架适合香港。
An Innovation, Technology and Industry Bureau spokesman said the government would draw on guidelines from other jurisdictions and consult a broad range of stakeholders to ensure its framework suited Hong Kong.
这些指引将为各行业提供实用指导,帮助机构识别、预防和管理来自人工智能代理的风险。它们还将帮助机构建立与其应用性质和潜在影响相称的保障措施。
The guidelines would provide practical guidance across sectors, helping organisations identify, prevent and manage risks from AI agents. They would also help organisations establish safeguards proportionate to the nature and potential impact of their applications.
法律责任的空白 麦先生表示,香港的法律责任通常取决于诸如谁与受害者有合同关系、是否有人疏忽以及是否存在犯罪意图等因素。但人工智能可能会暴露该框架中的空白。
Gaps in legal liability Mak said liability in Hong Kong would normally depend on factors such as who had a contract with the victim, whether anyone had been negligent and whether there had been criminal intent. But AI could expose gaps in that framework. “AI isn’t a legal person, so AI itself can never be liable,” he said.
“他表示,‘AI 不是法律主体,因此 AI 本身永远无法承担责任。’‘在某些情况下,存在没有任何人也承担责任的空白。链条中的每个人都尽到了合理注意。没有刑事意图。但 AI 代理人仍然能够造成伤害。’Mak 说,他预计司法部工作组将识别出这些空白。但他表示,城市也需要解决涉及海外 AI 模型的跨境争端问题,在香港强制证人出庭、获取证据以及证明因果关系方面可能会遇到困难。
“In some cases, there’s a gap where no human is also liable. Everyone has reasonable care in the chain. There’s no criminal intent. But still, the AI agent is able to cause harm.”Mak said he expected the Department of Justice’s working group to identify such gaps. But he said the city also needed to address cross-border disputes involving overseas AI models, where compelling witnesses to appear in Hong Kong, obtaining evidence and proving causation could be difficult.
他建议香港应探索保险安排,可能包括强制性投保,以便受害者在责任得到解决之前就能获得赔偿。但他警告,强制性投保可能会抑制投资,且目前仍不清楚该系统将如何运作。
He said Hong Kong should explore insurance arrangements, potentially including mandatory cover, that would allow victims to receive compensation before liability was resolved. But he cautioned that mandatory insurance could deter investment and that it remained unclear how such a system would work.
司法部发言人称,已成立跨部门工作组,以审查 AI 产品引发的事故法律责任。各司法管辖部门已基本完成对其政策领域 600 多部法规的初步审查。
A Department of Justice spokesman said a cross-departmental working group had been established to examine legal liability for accidents caused by AI products. Bureaus and departments had largely completed a preliminary review of more than 600 ordinances within their policy areas.
他表示,工作组将很快开始审查现行法律是否充分处理 AI 产品引发的事故或损害的责任问题,也将考虑立法和监管应对措施,包括指导原则。
He said the working group would shortly begin examining whether existing laws adequately addressed liability for accidents or damage caused by AI products. It would also consider legislative and regulatory responses, including guidelines.
关于 AI 产品使用的强制性第三方责任保险,该局发言人表示,将在部门审查结果及 AI 技术及其应用的发展基础上,考虑相应的政策措施。
Asked about mandatory third-party liability insurance for the use of AI products, the bureau spokesman said it would consider policy measures in light of the department’s review and developments in AI technology and its applications.