AI数据泄露与影子AI：2025年英国组织面临的法律雷区

Camilo Artiga-Purcell, General Counsel at Kiteworks, identifies some of the ever-increasing risks and potential consequences of rushing to use AI in legal practice
Kiteworks的总法律顾问卡米洛·阿蒂加-珀塞尔指出了在法律实践中急于使用人工智能所带来的一些日益增加的风险和潜在后果。
 
Picture a partner at a leading UK law firm, racing to finalise a high-stakes merger. With a deadline looming, they turn to a free online AI tool, uploading sensitive deal documents for rapid analysis. The tool delivers, and the work is completed on time. Months later, a rival firm using the same AI platform receives uncannily precise insights about the merger’s structure in an AI-generated response. An investigation reveals that the original documents were incorporated into the AI’s training data, inadvertently exposing confidential strategies. The fallout is swift: a regulatory probe, eroded client trust, and a legal battle over compromised attorney-client privilege.
想象一下，英国一家顶尖律师事务所的合伙人正在争分夺秒地敲定一项高风险的并购交易。由于截止日期临近，他们转而使用一款免费的在线人工智能工具，上传敏感的交易文件以进行快速分析。该工具完成了任务，工作得以按时完成。几个月后，另一家使用同一人工智能平台的竞争律所，在人工智能生成的回复中收到了关于该并购结构异常精准的见解。调查显示，原始文件已被纳入人工智能的训练数据，无意中泄露了机密策略。后果接踵而至：监管调查、客户信任受损，以及一场围绕受损的律师-客户特权的法律纠纷。
 
This scenario is not a hypothetical – it reflects a growing crisis across UK organisations. Legal departments and businesses are embracing artificial intelligence at an unprecedented rate, driven by its promise of efficiency in tasks like contract drafting and legal research. Yet, a survey of 300 corporate legal departments found that 81% are using unapproved AI tools without data controls, creating a legal and compliance minefield. For UK organisations, governed by the UK GDPR and facing emerging AI regulations, the risks are acute. Without action, legal teams face breaches of confidentiality, multimillion-pound fines, and reputational damage. This article explores the scale of this problem, its legal implications, and practical steps to safeguard sensitive data while leveraging AI responsibly.
这种情况并非假设——它反映了英国各组织正面临的一个日益严重的危机。在人工智能有望提高合同起草和法律研究等任务效率的推动下，法律部门和企业正以前所未有的速度采用人工智能。然而，一项针对300个企业法律部门的调查发现，81%的部门在使用未经批准的人工智能工具，且没有数据管控措施，这造成了一个法律和合规方面的雷区。对于受英国《通用数据保护条例》约束且面临新兴人工智能法规的英国组织来说，相关风险十分严峻。若不采取行动，法律团队可能会面临机密泄露、数百万英镑的罚款以及声誉受损等问题。本文探讨了这一问题的严重程度、其法律影响，以及在负责任地利用人工智能的同时保护敏感数据的切实步骤。
 
Scale of the Problem 问题的严重性
 
The adoption of AI in UK legal departments is surging, with tools promising to streamline contract reviews, legal research, and document analysis. However, this enthusiasm has birthed a dangerous trend known as “Shadow AI,” where employees use personal or unapproved AI tools for work tasks without oversight. According to a recent survey, 83% of in-house counsel use AI tools not provided by their organisations, and 47% operate without any governance policies. The Stanford AI Index Report highlights a 56% rise in AI-related incidents globally, with data leaks a primary concern. In the UK, 57% of organisations admit they cannot track sensitive data exchanges involving AI, amplifying the risk of breaches.
英国法律部门对人工智能的采用率正大幅上升，相关工具有望简化合同审查、法律研究和文档分析工作。然而，这种热情催生了一种被称为“影子人工智能”的危险趋势，即员工在未经监督的情况下使用个人或未经批准的人工智能工具处理工作任务。最近的一项调查显示，83%的内部法律顾问使用并非由其所在组织提供的人工智能工具，47%的人在没有任何治理政策的情况下使用这些工具。《斯坦福人工智能指数报告》强调，全球与人工智能相关的事件增加了56%，数据泄露是主要关注点。在英国，57%的组织承认他们无法追踪涉及人工智能的敏感数据交换，这加剧了数据泄露的风险。
 
We recently surveyed 461 organisations on this issue, across a range of industries, and the results reinforce these concerns with alarming specificity. Only 17% have automated controls with data loss prevention capabilities to block unauthorised AI access though the legal sector fares even worse, with just 15% implementing technical controls – the lowest of any industry surveyed. Perhaps most troubling for UK law firms, 38% of legal organisations admit that over 16% of data sent to AI tools contains private or sensitive information, with 23% reporting that more than 30% of their AI-processed data is private.
我们最近就这一问题对461家不同行业的机构进行了调查，结果以惊人的具体性印证了这些担忧。只有17%的机构拥有具备数据防泄漏功能的自动化控制措施，以阻止对人工智能的未授权访问，而法律行业的情况更糟，仅有15%的机构实施了技术控制措施——这在所有接受调查的行业中是最低的。或许对英国律师事务所而言最令人担忧的是，38%的法律机构承认，发送给人工智能工具的数据中，超过16%包含私人或敏感信息，23%的机构表示，其经人工智能处理的数据中，超过30%是私人信息。
 
The UK’s regulatory landscape heightens these challenges. The UK GDPR, aligned with the EU’s GDPR, imposes stringent obligations on data processing, storage, and cross-border transfers, with fines up to £17.5 million or 4% of global annual turnover for violations. The proposed UK AI Bill signals increased scrutiny of AI governance, while existing regulations like the Network and Information Systems (NIS2) Directive demand robust cybersecurity. For legal departments, a single employee uploading client data to an unapproved AI tool can expose privileged communications, trade secrets, or merger strategies to servers in unknown jurisdictions, undermining the foundations of legal practice.
英国的监管环境加剧了这些挑战。与欧盟《通用数据保护条例》（GDPR）保持一致的英国《通用数据保护条例》（UK GDPR），对数据处理、存储和跨境传输施加了严格的义务，违规者将面临最高1750万英镑或全球年营业额4%的罚款。拟议的英国《人工智能法案》预示着对人工智能治理的审查将加强，而《网络与信息系统（NIS2）指令》等现有法规则要求健全的网络安全措施。对于法律部门而言，哪怕有一名员工将客户数据上传至未经批准的人工智能工具，都可能导致特权通信、商业秘密或并购策略暴露在未知司法管辖区的服务器上，从而动摇法律执业的根基。
 
Legal and Compliance Risks 法律与合规风险
 
The legal and compliance risks of ungoverned AI use are profound for UK organisations. Data protection violations top the list. The UK GDPR requires organisations to establish a lawful basis for processing personal data, adhere to data minimisation principles, and ensure security by design. When lawyers upload client data to consumer AI tools like ChatGPT or Claude, they relinquish control over that information. The data may be processed via third-party APIs, stored on servers in multiple jurisdictions, or used to train AI models, all potentially breaching UK GDPR requirements. Such violations can trigger severe penalties and lasting reputational harm.
对于英国组织而言，不受管控地使用人工智能所带来的法律和合规风险是深远的。数据保护违规位居榜首。《英国通用数据保护条例》要求组织为处理个人数据确立合法依据，遵守数据最小化原则，并确保设计层面的安全性。当律师将客户数据上传至ChatGPT或Claude等消费级人工智能工具时，他们就失去了对这些信息的控制权。这些数据可能通过第三方应用程序接口进行处理，存储在多个司法管辖区的服务器上，或被用于训练人工智能模型，所有这些都可能违反《英国通用数据保护条例》的要求。此类违规行为可能引发严厉的处罚和持久的声誉损害。
 
Confidentiality and privilege concerns are equally grave. Attorney-client privilege, a bedrock of legal practice, can be waived when communications are shared with third-party AI providers. Consider when a UK litigation team uploaded privileged strategies to an AI tool, only to have opposing counsel argue successfully that privilege was lost, rendering years of communications discoverable. Trade secrets and intellectual property face similar risks, as AI platforms may inadvertently expose proprietary information through model outputs or data breaches, violating confidentiality agreements.
保密性和特权方面的担忧同样严重。律师-客户特权作为法律实务的基石，在与第三方人工智能提供商共享通信内容时可能会被放弃。试想一下，英国一个诉讼团队将享有特权的策略上传到某个人工智能工具后，对方律师成功辩称该特权已丧失，导致多年的通信内容都可能被披露。商业秘密和知识产权也面临类似风险，因为人工智能平台可能会通过模型输出或数据泄露无意中泄露专有信息，从而违反保密协议。
 
Regulatory compliance failures add further complexity. The NIS2 Directive mandates robust cybersecurity controls, while the Financial Conduct Authority (FCA) requires strict data governance for financial services firms. The Solicitors Regulation Authority (SRA) imposes ethical obligations under Rule 2.1, requiring solicitors to maintain competence in the technologies they use. Failure to understand AI risks can lead to disciplinary action, as seen in recent SRA investigations into tech mismanagement, where firms faced fines and reputational damage for inadequate data security. As AI regulations evolve, legal departments that fail to govern AI use risk becoming targets for enforcement actions. In this case, “Attorney-client privilege can be lost with a single upload.”
监管合规失败进一步增加了复杂性。《网络与信息系统安全指令2》（NIS2）要求实施强有力的网络安全控制，而英国金融行为监管局（FCA）则要求金融服务公司实施严格的数据治理。英国律师监管局（SRA）根据第2.1条规则规定了道德义务，要求律师对其使用的技术保持专业能力。对人工智能风险的不了解可能导致纪律处分，最近SRA对技术管理不善的调查就体现了这一点，在这些调查中，公司因数据安全不足而面临罚款和声誉损失。随着人工智能法规的不断发展，未能对人工智能使用进行治理的法律部门可能会成为执法行动的目标。在这种情况下，“一次上传就可能失去律师-客户特权。”
 
How AI Data Leaks Occur 人工智能数据泄露如何发生
 
AI data leaks stem from a mix of technical vulnerabilities and human error. When lawyers upload documents to consumer AI tools, the data may be used to train the AI model, stored indefinitely on external servers, or shared with third-party APIs without transparency. These platforms, not designed for the rigorous security needs of legal work, make it nearly impossible to retrieve or delete data once uploaded – a risk coined the “irrevocability problem.” This is particularly alarming for legal departments handling privileged or sensitive information.
人工智能数据泄露源于技术漏洞和人为失误的共同作用。当律师将文件上传到消费级人工智能工具时，这些数据可能会被用于训练人工智能模型、在外部服务器上无限期存储，或者在缺乏透明度的情况下与第三方应用程序接口共享。这些平台并非为满足法律工作严苛的安全需求而设计，一旦数据上传，几乎无法检索或删除——这种风险被称为“不可撤销问题”。对于处理特权信息或敏感信息的法律部门而言，这一点尤其令人担忧。
 
Common scenarios include lawyers using AI for contract drafting, legal research, or document analysis under tight deadlines. A junior associate might paste a draft settlement agreement into an unapproved AI tool to refine its language, unaware that the data is now stored on a server abroad. Similarly, a senior lawyer might use AI to summarise merger documents, not realising that the tool’s outputs could later reveal confidential strategies to client competitors, targets, or potential buyers. These actions, driven by the need for efficiency, create vulnerabilities that can lead to data leaks, regulatory violations, loss of privilege, or loss of bonafide competitive advantage.
常见场景包括律师在紧迫的截止日期下使用人工智能进行合同起草、法律研究或文档分析。一位初级律师可能会将一份和解协议草案粘贴到未经批准的人工智能工具中以优化其措辞，却不知道这些数据现在存储在国外的服务器上。同样，一位高级律师可能会使用人工智能来总结并购文件，却没有意识到该工具的输出后来可能会向客户的竞争对手、目标公司或潜在买家泄露机密策略。这些出于提高效率需求的行为会产生漏洞，可能导致数据泄露、违反法规、特权丧失或真正竞争优势的丧失。
 
Our recent survey cited above confirmed that these scenarios reflect current industry realities. Despite the legal profession’s heightened awareness of data risks – 31% of legal firms cite data leaks as their top AI concern, the highest of any sector – this awareness hasn’t translated into action: 15% of legal organisations operate with no formal AI data policies whatsoever, while 70% rely solely on human-dependent controls like training sessions and warning emails. This creates what the report calls an “awareness-action gap,” where firms recognise the danger but fail to implement the technical safeguards necessary to prevent catastrophic breaches.
我们上述提到的近期调查证实，这些场景反映了当前行业的实际情况。尽管法律行业对数据风险的意识有所提高——31%的律师事务所将数据泄露列为他们对人工智能的首要担忧，这一比例在所有行业中最高——但这种意识并未转化为行动：15%的法律机构完全没有正式的人工智能数据政策，70%的机构仅依赖于培训课程和警示邮件等依赖人工的管控措施。这就造成了报告中所说的“意识-行动差距”，即律所认识到了危险，却未能实施必要的技术防护措施来防止灾难性的数据泄露。
 
Real-World Scenarios 现实场景
 
The dangers of AI data leaks become clear when we imagine what could go wrong. Picture the scenario from our opening: a legal team uploads confidential merger documents to an AI tool for analysis. The platform uses those documents to train its model, and suddenly, sensitive deal information surfaces elsewhere − triggering expensive disputes and destroying client relationships.
当我们想象可能出现的问题时，人工智能数据泄露的危险就变得清晰起来。想象一下我们开篇提到的场景：一个法律团队将机密的合并文件上传到人工智能工具进行分析。该平台使用这些文件来训练其模型，突然间，敏感的交易信息出现在其他地方——引发昂贵的纠纷并破坏客户关系。
 
Consider another possibility: a UK company’s legal department runs personal data through an unauthorised AI tool. The result? A full GDPR investigation, hefty fines, and damaging headlines that tarnish the firm’s reputation. Perhaps most alarming is this scenario: a litigation team uploads privileged attorney-client communications to an AI platform. When opposing counsel discovers this, they successfully argue that privilege has been waived. The entire case strategy unravels, and what should have been protected conversations become fair game in court.
再考虑另一种可能性：一家英国公司的法律部门通过未经授权的人工智能工具处理个人数据。结果会怎样？全面的《通用数据保护条例》调查、高额罚款，以及有损公司声誉的负面新闻标题。或许最令人担忧的是这种情况：一个诉讼团队将享有特权的律师与客户的通信上传到人工智能平台。当对方律师发现这一点时，他们会成功辩称特权已被放弃。整个案件策略会瓦解，本应受到保护的对话在法庭上会变成公开可利用的信息。
 
These aren’t just theoretical risks, they represent the very real consequences that await organisations operating without proper AI governance. Each scenario shows how quickly simple upload can transform into a professional catastrophe.
这些并非只是理论上的风险，它们代表着那些缺乏适当人工智能治理的组织将面临的非常真实的后果。每个场景都展示了一次简单的上传会多么迅速地演变成一场重大的职业灾难。
 
Building a Compliant AI Framework 构建合规的人工智能框架
 
To mitigate these risks, UK legal departments must establish a robust AI governance framework tailored to their needs. The foundation is a clear governance structure. Comprehensive AI usage policies should outline acceptable tools, data handling protocols, and consequences for non-compliance, addressing confidentiality, privilege, and data security. Regular risk assessments are vital to identify vulnerabilities, while a formal approval process ensures only secure, compliant AI platforms are used.
为了缓解这些风险，英国法律部门必须建立一个符合自身需求的强大人工智能治理框架。其基础是清晰的治理结构。全面的人工智能使用政策应明确可接受的工具、数据处理协议以及违规后果，并解决保密性、特权和数据安全问题。定期风险评估对于识别漏洞至关重要，而正式的审批流程则确保只使用安全、合规的人工智能平台。
 
Technical controls are critical. Data classification systems should identify sensitive information before it is processed by AI tools. Access controls, such as role-based permissions and monitoring, can prevent unauthorised use of consumer AI platforms. An approved list of enterprise-grade AI tools, designed with legal and compliance requirements in mind, ensures efficiency without sacrificing security. These tools must integrate with existing cybersecurity infrastructure and incorporate data loss prevention measures to protect sensitive information.
技术控制至关重要。数据分类系统应在敏感信息被人工智能工具处理之前对其进行识别。访问控制（如基于角色的权限和监控）可以防止消费者人工智能平台被未授权使用。一份经过批准的企业级人工智能工具清单，在设计时考虑了法律和合规要求，能够在不牺牲安全性的前提下确保效率。这些工具必须与现有的网络安全基础设施相集成，并纳入数据防泄漏措施，以保护敏感信息。
 
Training and awareness underpin effective governance. Mandatory training for all legal staff, from partners to associates, should cover the technical and legal risks of AI, including UK GDPR obligations and SRA requirements. Regular updates on emerging threats, such as new data breach tactics or regulatory changes, keep teams informed. Clear reporting mechanisms for AI-related incidents foster transparency and enable swift responses to potential breaches, minimising damage.
培训和意识是有效治理的基础。所有法律人员（从合伙人到律师助理）都必须接受培训，内容应涵盖人工智能的技术和法律风险，包括英国通用数据保护条例的义务和律师监管局的要求。定期更新新出现的威胁（如新的数据泄露策略或法规变化），能让团队了解相关情况。明确的人工智能相关事件报告机制有助于提高透明度，并能对潜在的违规行为做出快速反应，从而将损害降到最低。
 
Practical Recommendations for Legal Teams
给法律团队的实用建议
 
Legal teams must act swiftly to address AI data risks, with immediate, medium-term, and long-term strategies. In the short term, conducting a Shadow AI audit is essential to uncover unapproved tool usage. This involves surveying staff to identify all AI tools in use, assessing the data being processed, and documenting potential exposures. This could be backed up by technical solutions, such as an “AI Gateway” to help enforce these policies by automatically detecting and blocking sensitive client data from reaching unauthorised AI platforms, providing real-time protection while policies are developed. Emergency controls, such as blocking access to consumer AI platforms and providing approved alternatives, can halt further risks. Clear communication ensures staff understand the urgency and comply with new protocols.
法律团队必须迅速采取行动，应对人工智能数据风险，并制定即时、中期和长期策略。短期内，开展“影子人工智能”审计至关重要，以发现未经批准的工具使用情况。这包括调查员工，确定所有正在使用的人工智能工具，评估正在处理的数据，并记录潜在的风险暴露点。这可以通过技术解决方案来支持，例如“人工智能网关”，通过自动检测和阻止敏感客户数据流向未授权的人工智能平台来帮助执行这些政策，在政策制定过程中提供实时保护。紧急控制措施，如阻止访问消费者人工智能平台并提供经批准的替代方案，可以遏制进一步的风险。清晰的沟通能确保员工理解事情的紧迫性并遵守新协议。
 
In the medium term, comprehensive AI policies should align with UK GDPR, SRA, and FCA requirements. Again technical controls, not just documentation, could be used to apply sensitive data classification, access controls and audit trails, regardless of which AI tool employees attempt to use. Vendor vetting procedures are crucial, ensuring AI providers meet stringent security and compliance standards, with contracts that protect client data and include audit rights. An AI-specific incident response plan prepares teams to act decisively in case of a breach, minimising regulatory and reputational fallout.
从中期来看，全面的人工智能政策应符合英国《通用数据保护条例》、律师监管局（SRA）和金融行为监管局（FCA）的要求。同样，无论员工尝试使用哪种人工智能工具，都可以采用技术控制（而非仅仅依靠文件记录）来实施敏感数据分类、访问控制和审计跟踪。供应商审查程序至关重要，要确保人工智能提供商符合严格的安全和合规标准，同时签订的合同需保护客户数据并包含审计权。专门针对人工智能的事件响应计划能让团队在发生数据泄露时果断采取行动，将监管风险和声誉损失降至最低。
 
For the long term, investing in enterprise-grade AI solutions designed for legal work, such an AI Gateway described above, is vital. Annual policy reviews keep governance measures aligned with evolving technology and regulations, embedding AI governance into the broader compliance strategy to maintain client trust while leveraging AI’s benefits.
从长远来看，投资专为法律工作设计的企业级人工智能解决方案（例如上述的人工智能网关）至关重要。年度政策审查使治理措施与不断发展的技术和法规保持一致，将人工智能治理融入更广泛的合规战略中，从而在利用人工智能优势的同时维护客户信任。
 
Future Outlook and Conclusion 未来展望与结论
 
The UK’s regulatory landscape is evolving rapidly, with the proposed AI Bill, UK GDPR, and NIS2 Directive signalling heightened scrutiny of AI governance. Legal departments that fail to act risk becoming cautionary tales, facing fines, client loss, and reputational damage. Conversely, those that implement robust governance can gain a competitive edge, demonstrating to clients their commitment to security and compliance while harnessing AI’s efficiency.
英国的监管环境正在迅速演变，拟议的《人工智能法案》、英国《通用数据保护条例》和《网络与信息系统安全指令2》表明对人工智能治理的审查将更加严格。未能采取行动的法律部门可能会成为警示案例，面临罚款、客户流失和声誉损害。相反，那些实施健全治理的部门可以获得竞争优势，向客户展示其对安全和合规的承诺，同时利用人工智能的效率。
 
The urgency of addressing AI data leaks is undeniable. Legal teams must act now to audit AI usage, implement controls, and educate staff. By balancing innovation with risk management, UK organisations can protect sensitive data, uphold client trust, and navigate a complex regulatory landscape. The legal profession is built on trust and diligence. In the AI era, these principles demand proactive governance to ensure technology serves as a tool for progress, not a source of peril.
解决人工智能数据泄露问题的紧迫性毋庸置疑。法律团队必须立即采取行动，审核人工智能的使用情况，实施管控措施，并对员工进行培训。通过在创新与风险管理之间取得平衡，英国的组织能够保护敏感数据，维护客户信任，并应对复杂的监管环境。法律行业建立在信任和勤勉的基础之上。在人工智能时代，这些原则要求我们进行前瞻性治理，以确保技术成为推动进步的工具，而非危险的源头。
 
Author :Camilo Artiga-Purcell serves as General Counsel at Kiteworks, where he leads legal strategy and governance initiatives for secure content communications and collaboration. With extensive experience in data privacy, cybersecurity, and emerging technology law, he advises organizations on managing AI-related risks while maintaining competitive advantage.
作者：卡米洛·阿蒂加-珀塞尔担任Kiteworks的总法律顾问，负责领导安全内容通信和协作方面的法律战略及治理计划。凭借在数据隐私、网络安全和新兴技术法方面的丰富经验，他为各组织提供有关管理人工智能相关风险同时保持竞争优势的建议。