AI Security Institute

The AI Security Institute is a research organisation within the United Kingdom's Office for the Prime Minister and Cabinet, aiming to equip governments with a scientific understanding of the risks posed by advanced artificial intelligence.人工智能安全研究所是英国首相与内阁办公室下属的研究机构，目标是让各国政府具备对先进人工智能所带来风险的科学认识。

It conducts research and develops and tests mitigations, was the first institution of its kind, and has since seen similar AI safety institutions modelled after it established by governments of other countries.它开展研究并开发与检验风险缓解措施；它是此类机构中的第一家，此后其他国家的政府也参照它设立了类似的人工智能安全机构。

History

• The organisation was initially named the AI Safety Institute and was launched in November 2023 under the Rishi Sunak premiership.该机构最初名为「人工智能安全研究所」，于 2023 年 11 月在里希·苏纳克任首相期间启动。
• It was launched as part of the AI Safety Summit 2023 event at Bletchley Park.它的启动是布莱切利园「2023 人工智能安全峰会」活动的一部分。
• Sunak had announced the plan to establish the institute the week before, during a press conference on 26 October 2023.苏纳克在此前一周、即 2023 年 10 月 26 日的记者会上宣布了设立该机构的计划。
• In February 2025, under the premiership of Keir Starmer, the institute was renamed the AI Security Institute.2025 年 2 月，在基尔·斯塔默任首相期间，该机构更名为「人工智能安全研究所」。
• The change was announced by the Secretary of State for Science, Innovation and Technology, Peter Kyle, at the Munich Security Conference.这一更名由科学、创新与技术大臣 Peter Kyle 在慕尼黑安全会议上宣布。
• It was intended to reflect a renewed focus on serious AI risks with security implications.此举意在体现其重新聚焦于具有安全影响的严重人工智能风险。
• Those risks include how the technology can be used to develop chemical and biological weapons, how it can be used to carry out cyber-attacks, and how it can enable crimes such as fraud and child sexual abuse.这些风险包括该技术如何被用于研制化学与生物武器、如何被用于发动网络攻击，以及如何为欺诈与儿童性虐待等犯罪提供便利。
• In July 2026 the Department for Science, Innovation and Technology was dissolved as part of a broader governmental reorganisation.2026 年 7 月，科学、创新与技术部作为更广泛的政府重组的一部分被撤销。
• As a result of that reorganisation the institute was relocated under the newly formed Office for the Prime Minister and Cabinet.作为这次重组的结果，该机构被划归新设立的首相与内阁办公室。
• Because it was established at a summit and restructured twice within three years, its organisational history reflects how quickly government attention to AI risk has shifted.由于它在一场峰会上设立，并在三年内两度重组，其机构史反映出政府对人工智能风险的关注变化之快。

Mission and activities

• Its mission is to equip governments with a scientific understanding of the risks posed by advanced AI.其使命是让各国政府具备对先进人工智能所带来风险的科学认识。
• It conducts technical research on the risks AI poses to national security and public safety.它就人工智能对国家安全与公共安全构成的风险开展技术研究。
• It monitors the fast-moving landscape of AI development.它监测快速演变的人工智能发展格局。
• It advances solutions such as safeguards, alignment and control.它推进保障措施、对齐与控制等解决方案。
• It also develops and tests risk mitigations rather than only describing risks.它还开发并检验风险缓解措施，而不仅是描述风险。
• It works with AI developers to encourage responsible development, and informs policymakers about current and emerging risks.它与人工智能开发者合作以鼓励负责任的开发，并就当前与新兴风险向决策者通报。
• It is described as designed like a startup in government, combining the authority of government with the expertise and agility of the private sector.它被描述为「像政府中的初创企业」，把政府的权威与私营部门的专长和敏捷性结合起来。
• It is overseen by the Minister of Artificial Intelligence, a post held by Kanishka Narayan.它由人工智能事务大臣监督，该职由 Kanishka Narayan 担任。
• As of February 2026 it had access agreements with Anthropic, Google and OpenAI to test their models before release.截至 2026 年 2 月，它与 Anthropic、谷歌与 OpenAI 签订了访问协议，可在其模型发布前进行测试。
• It maintains an open-source platform called Inspect, which permits companies, governments and academics to run standardised safety tests for AI usage.它维护一个名为 Inspect 的开源平台，使企业、政府与学术界能够对人工智能的使用开展标准化安全测试。
• Among its work is the reported detection of multiple serious vulnerabilities that could enable the development of biological weapons.其工作之一是被报道发现了多个可能被用于研制生物武器的严重漏洞。
• Those vulnerabilities were fixed before the relevant model was launched.这些漏洞在相关模型发布之前已得到修复。
• It has published work on the persuasive capabilities of conversational AI following large-scale experiments.在开展大规模实验之后，它发表了关于对话式人工智能说服能力的研究。
• One study found that large language models post-trained for political persuasiveness became systematically less accurate and up to fifty-one per cent more persuasive on political issues.一项研究发现，为政治说服力进行后训练的大型语言模型在政治议题上系统性地变得更不准确，说服力最高提升 51%。
• It has also worked on the use of AI for emotional needs.它还研究了人工智能在满足情感需求方面的使用。
• In December 2025 it found that nearly ten per cent of United Kingdom citizens used systems such as chatbots for emotional purposes on a weekly basis.2025 年 12 月，它发现近 10% 的英国公民每周都会出于情感目的使用聊天机器人等系统。
• At the same time it found that systems were outperforming PhD-level researchers on scientific knowledge tests and helping non-experts succeed at laboratory work previously out of reach.同时它发现，这些系统在科学知识测试中已超过博士级研究者，并帮助非专家完成以往无法企及的实验室工作。
• It has also evaluated the cyber capabilities of leading open and closed weight models, comparing how far behind the frontier open models run.它还评估了领先的开源与闭源模型的网络攻防能力，比较开源模型落后前沿多远。
• It works with the wider research community, AI developers and other governments to affect how AI is developed and to shape global policymaking.它与更广泛的研究界、人工智能开发者以及其他国家政府合作，以影响人工智能的开发方式并塑造全球政策制定。
• Its published outputs include a Frontier AI Trends Report alongside its research papers.其公开成果包括《前沿人工智能趋势报告》以及各类研究论文。
• It publishes a blog and lists grants as further channels for its work.它还发布博客并列出资助机会，作为其工作的其他渠道。

Organization and governance

• The institute is part of the United Kingdom's Office for the Prime Minister and Cabinet.该机构隶属英国首相与内阁办公室。
• It is overseen by the Minister of Artificial Intelligence.它由人工智能事务大臣监督。
• It was previously within the Department for Science, Innovation and Technology, until that department was dissolved in 2026.它此前隶属于科学、创新与技术部，直至该部于 2026 年撤销。
• It is designed to combine the authority of government with the expertise and agility of the private sector.它的设计要点在于结合政府的权威与私营部门的专长与敏捷性。
• Its work is organised into technical research and what it describes as global impact, the latter covering cooperation with developers and informing policymakers.其工作分为技术研究与它所称的「全球影响」两部分，后者涵盖与开发者合作及向决策者通报。
• It runs an open-source toolchain, allowing external parties to reproduce its standardised tests rather than relying on internal assessment alone.它运营开源工具链，使外部各方能够复现其标准化测试，而不必仅依赖内部评估。
• Its access agreements with model developers give it pre-release testing rights that few other bodies hold.它与模型开发者签订的访问协议赋予其发布前测试权限，而拥有此类权限的机构并不多。

Membership and funding

• The institute is state-backed and does not operate on membership dues.该机构由国家支持，不依靠会员会费运作。
• It was the first state-backed organisation dedicated to understanding and mitigating advanced AI risks.它是首家致力于理解并缓解先进人工智能风险的国家支持机构。
• Its funding comes from government, reflecting the view that governments have a critical role in ensuring advanced AI is safe, secure and beneficial.其经费来自政府，反映出一种判断：政府在确保先进人工智能安全、可靠且有益方面负有关键作用。
• Its access arrangements with developers are negotiated rather than purchased, giving it testing access without owning the models.它与开发者之间的访问安排通过协商而非购买获得，使其无需拥有模型即可开展测试。
• It lists grants as part of its work, indicating support for external research alongside its in-house programme.它把资助列为其工作的一部分，表明在自有项目之外也支持外部研究。
• Because it was the first institution of its kind, other governments have modelled similar AI safety institutions on it rather than starting from scratch.由于它是此类机构中的第一家，其他政府参照它设立类似的人工智能安全机构，而非从零起步。

Reception

• The institute was the first institution of its kind, and similar bodies in other countries have been modelled after it.该机构是此类机构中的第一家，其他国家的类似机构是参照它设立的。
• Its launch at the Bletchley Park summit tied its creation to a moment of international attention to frontier AI risk.它在布莱切利园峰会上的启动，使其设立与国际社会关注前沿人工智能风险的时刻相连。
• The 2025 renaming is often cited as evidence that government attention shifted from general safety concerns to security-specific risks.2025 年的更名常被视为政府关注点从一般安全顾虑转向具体安全风险的证据。
• Its report of serious vulnerabilities with biological-weapon potential, fixed before release, illustrates pre-release testing in practice.其关于具生物武器潜力的严重漏洞在发布前被修复的报告，展示了发布前测试的实际作用。
• Its finding that models post-trained for political persuasiveness became less accurate is widely cited in debates about AI and political discourse.其关于为政治说服力后训练的模型变得更不准确的发现，在关于人工智能与政治话语的讨论中被广泛引用。
• The same research programme produced quantitative evidence about public use of chatbots for emotional purposes and about model performance on scientific knowledge.同一研究计划还就公众出于情感目的使用聊天机器人，以及模型在科学知识上的表现，给出了量化证据。
• Its open-source Inspect platform is notable for letting outside parties run the same tests the institute uses.其开源平台 Inspect 值得注意之处，在于允许外部各方运行与该机构相同的测试。
• Two departmental reorganisations within three years have raised questions about the stability of its institutional home.三年内两度部门重组，引发人们对其机构归属稳定性的疑问。

Notable people

• Rishi Sunak, as Prime Minister, announced and then launched the institute in 2023.里希·苏纳克在任首相期间于 2023 年宣布并启动该机构。
• Keir Starmer, as Prime Minister, oversaw the 2025 renaming from AI Safety Institute to AI Security Institute.基尔·斯塔默在任首相期间主持了 2025 年从「人工智能安全研究所」到「人工智能安全研究所（Security）」的更名。
• Peter Kyle, Secretary of State for Science, Innovation and Technology, announced the renaming at the Munich Security Conference.科学、创新与技术大臣 Peter Kyle 在慕尼黑安全会议上宣布了更名。
• Kanishka Narayan serves as the Minister of Artificial Intelligence overseeing the institute.Kanishka Narayan 担任监督该机构的人工智能事务大臣。
• Its research is carried out by in-house technical staff who also build and maintain its open testing platform.其研究由内部技术人员开展，他们同时也构建并维护其开放测试平台。
• Its access agreements bring it into regular contact with the safety teams of Anthropic, Google and OpenAI, and give it pre-release testing rights that few other bodies hold.其访问协议使它同 Anthropic、谷歌与 OpenAI 的安全团队保持经常性接触，并赋予其发布前测试权限，而持有此类权限的机构并不多。
• Its work covers both understanding what advanced models can do and developing the safeguards, alignment techniques and control measures that governments might need.其工作既包括了解先进模型能做什么，也包括开发政府可能需要的保障措施、对齐技术与控制手段。
• Because it publishes its testing platform openly, external researchers and other governments can reproduce its evaluations rather than accepting its conclusions on trust.由于它公开其测试平台，外部研究者与其他国家政府可复现其评估，而不必凭信任接受其结论，这一点对于跨国比较模型行为尤其重要。
• Its research on cyber capability compares open and closed weight models, which is intended to show how quickly publicly available models approach the frontier of capability, a question that matters for export control and for how quickly safeguards need updating.其关于网络攻防能力的研究比较开源与闭源模型，意在显示公开可得的模型逼近能力前沿的速度有多快，这一问题对出口管制以及保障措施需要多快更新都十分重要，也关系到各国如何判断开源模型发布的节奏。
• Its findings on persuasion and on emotional use of chatbots are intended to inform policymakers about risks that are not purely technical, and its published outputs include a Frontier AI Trends Report alongside its research papers, giving ministers and officials regular assessments rather than one-off findings on a fast-moving and only partly understood class of technology whose behaviour changes from one release to the next.其关于说服力以及聊天机器人情感化使用的发现，意在让决策者了解并非纯属技术性质的风险；其公开成果还包括《前沿人工智能趋势报告》以及各类研究论文，从而就一类快速演变且仅有部分被理解的新兴技术，为部长与官员提供定期评估而非一次性结论，便于政策随技术演进而调整。