[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f-BOgNBiWl7euNkGlb23dXFaPhDp_2XqZlh7OOQNH3Xk":3,"$fDuOfbrGklFg9DNstLQj01v77JKgm8YTNP3W1V3nd5U4":77,"faq-section-":233,"white_papers":234},{"tableOfContents":4,"markDownContent":5,"htmlContent":6,"metaTitle":7,"metaDescription":8,"wordCount":9,"readTime":10,"title":11,"nbDownloads":12,"excerpt":13,"lang":14,"url":15,"intro":8,"featured":4,"state":16,"author":17,"authorId":18,"datePublication":22,"dateCreation":23,"dateUpdate":24,"mainCategory":25,"categories":41,"metaDatas":66,"imageUrl":67,"imageThumbUrls":68,"id":76},false,"AI agents no longer just answer questions. They plan, call tools, keep a memory and act: they send emails, update records, screen candidates or trigger payments, sometimes before anyone reviews the result.\n\nFor privacy and compliance teams, this changes the question. It is no longer only who can use AI, but what an agent is allowed to reach, what it is allowed to do, and who answers for the result.\n\nTo this date, no law has been written specifically for AI agents. However, data protection authorities, cybersecurity agencies and public bodies in Europe, the UK and Asia have published their guidance on agentic AI, and international standards such as the NIST AI RMF and ISO/IEC 42001 provide a common method. Each covers part of the picture. Read together, they show a clear consensus on what organisations will be expected to do.\n\nThis article maps those frameworks: who published what, what each one focuses on, and where they agree.\n\n## Why AI agents need their own governance\n\nAgents raise the same legal questions as other AI systems, but they raise the stakes in three ways:\n\n- **Autonomy.** An agent decides how to reach a goal, so its actions can go beyond what the user intended.\n- **Access.** To be useful, agents often need broad rights over browsers, mailboxes, files, calendars and business tools.\n- **Action.** Agents have real-world effects, often before a human sees the file.\n\n## The risks regulators agree on\n\nRegulators use different words, but they describe the same set of risks.\n\n**1. Roles that shift within a single task.** In a classic vendor setup, it is fairly clear who decides and who acts on instructions. With agents, roles can change from one step to the next: the agent calls a third-party API, queries an external database or chains a sub-agent. A supplier described as a processor in the contract can become an independent or joint controller for part of the task. The more actors involved, the harder it is to say who answers for what.\n\n**2. Purpose drift.** Purpose limitation is tied to lawful basis: an agent that goes beyond its stated purpose is likely outside its lawful basis too. An agent asked to screen CVs for technical skills may also infer age from graduation dates or \"culture fit\" from social media. That creates new purposes, with no legal basis and no information to the candidate. The problem can also start upstream, when purposes are defined too broadly for open-ended, general-purpose agents.\n\n**3. Excessive access.** Agents often get broad default rights over files, emails, credentials and third-party accounts. Every source the agent can reach is personal data it may process, store or pass on. Excessive permissions also make the agent a target: anyone who compromises it inherits its access. And an action can be authorised and still cause harm, because an agent can make mistakes or combine permitted operations in a dangerous way.\n\n**4. Inferences and persistent memory.** Agents draw conclusions nobody asked for. Some of them fall into sensitive categories, such as health or religion, without any legal basis covering them. Agents also remember: their persistent memory can build a detailed profile of a person, spread across several agents. Memories, histories and logs become stores of personal data in their own right, each needing a retention period.\n\n**5. Errors that spread.** A generative tool that makes a mistake produces a wrong answer. An agent that makes a mistake can act on it, pass it to another agent and build on it. Inaccurate or hallucinated data can spread across a multi-agent system without alerting anyone, and the same request may not produce the same actions twice.\n\n**6. Oversight that is not effective.** A human reviewer does not make a system safe by itself. Reviewers trust outputs that have been reliable in the past, approve too quickly under time pressure, and lack the information or authority to say no. A click-through approval can leave a decision effectively automated.\n\n**7. A wider attack surface.** Agents read untrusted content, such as emails, documents and web pages, and can act on it. A prompt injection, a malicious instruction hidden in that content, can lead to real actions: data sent to an external server, records changed, messages sent. Plugins and connectors that have not been security-reviewed add further risk.\n\n**8. Harder rights and transparency.** The more agents, memories and third-party services involved, the harder it is to tell people how their data is used. It is also harder to answer an access or erasure request completely. Which agent collected the data, where it is stored, and which service received it are questions many organisations cannot yet answer.\n\n## The legal baseline: the EU example\n\nBefore looking at the guidance, it is worth remembering that the law already applies. Unlike the frameworks below, it is not voluntary.\n\n**Data protection law applies in full.** No jurisdiction has created a special regime for agents yet. However, agents are not legal persons. The organisation that decides why and how data is processed remains accountable, whether or not an agent does the processing. In Europe for example, this means the GDPR, including its rules on automated decision-making (Article 22).\n\n**The EU AI Act applies as law.** The AI Act does not define AI agents, but agents fall under its existing rules depending on how they are used:\n\n- **Transparency (Article 50)** has applied since 2 August 2026 to agents that interact with people or generate content.\n- **High-risk obligations** apply where an agent falls into a high-risk category, for example in recruitment or credit.\n- **Prohibited practices (Article 5)**, such as harmful manipulation or the exploitation of vulnerabilities, apply to agents like any other AI system.\n- **AI literacy (Article 4)** requires organisations to train the staff who use agents.\n\n## International standards: the method and the structure\n\nTwo international references are not specific to agents, but they give organisations a common way to govern them alongside the law.\n\n**The NIST AI Risk Management Framework (AI RMF)** Published by the US National Institute of Standards and Technology in January 2023, the AI RMF is voluntary guidance. It organises AI risk management around four functions:\n\n- **Govern:** set roles, policies and accountability;\n- **Map:** understand the context and the risks of each use case;\n- **Measure:** assess and test those risks;\n- **Manage:** treat them and monitor them over time.\n\nIn July 2024, NIST added a profile for generative AI (NIST AI 600-1). In 2026, its Center for AI Standards and Innovation (CAISI) launched an AI Agent Standards Initiative to develop standards specific to agents.\n\n**ISO/IEC 42001** Published in 2023 by ISO and IEC, ISO/IEC 42001 is the first certifiable standard for an AI management system. It works like ISO 27001 does for information security: policies, roles, risk assessment, controls, internal audits and continuous improvement. For agents, it gives a structure to keep the inventory, risk assessments and controls up to date as agents change.\n\n**How they fit together**\\\nIn practice, an organisation can use the NIST AI RMF to assess the risks of each agent, ISO/IEC 42001 to organise its governance, and the AI Act and data protection law as the requirements that must be met. \n\n## The regulatory frameworks at a glance\n\nThe frameworks below do not create new obligations. They show how regulators expect these rules to be applied in practice.\n\n| Framework | Publisher | Date | Main focus |\n| --- | --- | --- | --- |\n| Exploratory note on agentic AI | CNIL and CIANum (France) | July 2026 | Data protection risks, loss of user control |\n| *Securing an agentic AI system* | ANSSI (France) | October 2026 | Security principles for deploying agents |\n| Agentic AI from a data protection perspective | AEPD (Spain) | February 2026 | Detailed technical and GDPR analysis |\n| Tech Futures: Agentic AI, and call for evidence | ICO (United Kingdom) | January and October 2026 | Novel data protection risks |\n| TechDispatch on human oversight | EDPS (EU) | September 2025 | Limits of human-in-the-loop |\n| Model Personal Data Protection Framework for agentic AI | PCPD (Hong Kong) | August 2026 | Data protection principles and security checklist |\n| Model AI Governance Framework for Agentic AI | IMDA (Singapore) | v1.5, May 2026 | Operational governance, with case studies |\n\n## European Union\n\n### The EDPS: a human in the loop is not enough on its own\n\nThe European Data Protection Supervisor's TechDispatch on human oversight of automated decision-making challenges a common assumption: adding a human reviewer does not make a system safe by itself. The EDPS identifies flawed assumptions behind many oversight designs:\n\n- that the system will recognise its own limits and hand over to a human;\n- that the reviewer will not be influenced by the output (automation bias);\n- that the reviewer will always have the time, expertise and authority to override it.\n\nIts recommendations include enough review time, training, interfaces that make intervention easy, and regular audits of how reviewers decide.\n\n## France\n\n### The CNIL and CIANum: loss of control at the heart of the analysis\n\nPublished in July 2026 with the French AI and Digital Council, the CNIL's exploratory note describes an \"equation with multiple unknowns\" for users. Its main concerns:\n\n- **Persistent memory** that builds a deep, evolving profile of the person, spread across several agents.\n- **Autonomous execution** that may diverge from what the user intended.\n- **Unclear liability** when many actors are involved.\n- **Data spread across agents**, which makes rights to access, rectification and erasure harder to exercise.\n\nThe note is not binding. It confirms that the GDPR applies in full, and recommends four safeguards:\n\n- detecting and filtering misuse;\n- separating memory by agent and by processing;\n- keeping human oversight over critical actions;\n- a kill switch.\n\n### ANSSI: authorised does not mean safe\n\nFrance's national cybersecurity agency published *Securing an agentic AI system* on 30 October 2026. It concerns any organisation that deploys or integrates an agent, including companies that buy one off the shelf and connect it to their own tools.\n\nIts starting point: an action the agent is authorised to take can still cause harm, because an agent can make mistakes, be manipulated, or combine permitted operations in a dangerous way. Its seven principles cover:\n\n- end-to-end risk analysis;\n- limiting autonomy;\n- sorting actions into autonomous, subject to approval, or prohibited;\n- separating environments;\n- controls that do not depend on the model;\n- limiting the impact of the system;\n- demonstrating control over time.\n\nFor legal teams, the security requirements from the risk analysis should appear in the specifications of any product or service purchased.\n\n## Spain\n\n### The AEPD: a detailed analysis of risks specific to agents\n\nThe Spanish authority's guidance, published in February 2026, focuses on what is specific to agents beyond the risks already known from generative AI:\n\n- **Excessive permissions.** They turn agents into pivot points for attackers.\n- **Errors that compound** over long tasks.\n- **Memories and logs that become large stores of personal data.**\n- **\"BYOAgentic\".** Employees build their own agents outside any governance.\n\nThe AEPD also warns that relying entirely on users to catch errors is itself a threat. It recommends deciding the level of autonomy for each processing operation, from \"the agent proposes, the human acts\" to \"the agent acts, the human observes\".\n\n## United Kingdom\n\n### The ICO: eight novel data protection risks\n\nIn its Tech Futures report of January 2026, the ICO mapped the data protection risks specific to agentic AI:\n\n- unclear controller and processor roles across the supply chain;\n- more automated decision-making;\n- purposes set too broadly for open-ended agents;\n- processing beyond what is necessary;\n- unintended inference of special category data;\n- transparency and rights made harder by complexity;\n- new cyber threats;\n- the concentration of personal data in personal assistant agents.\n\nIn October 2026, the ICO opened a call for evidence on agentic AI, closing on 20 November 2026. It seeks evidence on several topics such as agent identities and permissions or meaningful human oversight.\n\nThe responses will feed into future guidance and the ICO's statutory code of practice on AI and automated decision-making.\n\n## Asia\n\n### The PCPD (Hong Kong): nine recommendations and a security checklist\n\nPublished in August 2026, the PCPD's framework applies Hong Kong's six Data Protection Principles to agents. Its premise is shared by every regulator: AI agents are not legal persons, and the organisation remains accountable.\n\nIt identifies five risks:\n\n- extensive default access;\n- system vulnerabilities;\n- unvetted plugins and skills;\n- function creep;\n- errors cascading between agents.\n\nIts nine recommendations range from avoiding data collected \"just in case\" to keeping a human in the loop for high-stakes or irreversible actions. A practical security checklist covers each stage of an agent's life, from evaluation to uninstallation.\n\n### IMDA (Singapore): the operational framework\n\nSingapore's Model AI Governance Framework for Agentic AI, now in version 1.5, is built around four dimensions:\n\n1. **Assess and bound the risks upfront.** Score each use case on impact and likelihood, then limit what the agent can reach.\n2. **Make humans meaningfully accountable.** Allocate responsibilities across the value chain and set approval checkpoints.\n3. **Implement technical controls and processes.** Test whole workflows, roll out gradually and monitor continuously.\n4. **Enable end-user responsibility.** Tell users what the agent can do, train staff, and protect the skills they need to spot errors.\n\nTwo ideas from IMDA are useful well beyond Singapore. The first distinguishes an agent's action-space (what it can do) from its autonomy (how freely it decides). The second is concrete measures against automation bias, such as tracking how often reviewers override the agent and how long they take to decide. The framework is illustrated with case studies from banks, technology companies and public bodies.\n\n## Where the frameworks agree\n\nAcross very different jurisdictions, the same expectations come back:\n\n1. **Accountability stays with the organisation.** Agents are not legal persons, and someone must answer, by name, for what each agent does.\n2. **Know your agents.** Every agent should be listed, with its owner, its purpose, its access and its risk level.\n3. **Scale controls to the risk.** Autonomy, access to sensitive data and the reversibility of actions are the main factors.\n4. **Limit access to what the task needs.** Enforce it in the system, not only through instructions in the prompt.\n5. **Bound memory and inferences.** Set retention periods, filter sensitive inferences, and make sure rights work across every memory.\n6. **Make human oversight real.** Put checkpoints on high-stakes actions, and give reviewers the time and authority to say no.\n7. **Trace and monitor.** Keep logs that reconstruct each task, test before go-live, and reassess after every change.\n\n## What it means for privacy and compliance teams\n\nThese frameworks are not binding. But they show how existing rules, starting with the GDPR and the AI Act, will be read and enforced when agents are involved. Organisations that wait for agent-specific legislation are taking a risk.\n\nThree first steps to take:\n\n1. **Find out which agents are already running.** Include features switched on in existing software and agents built by employees.\n2. **Check your existing governance.** Make sure your policies, AI register, contracts and impact assessments cover what agents actually do.\n3. **Assess before deploying.** Do it at the start of each project, not after go-live.\n\n**Go further** Dastra helps privacy and AI compliance teams govern their AI agents in one place: build your AI register, assess suppliers and use cases, track risks and measures, and connect everything to your GDPR compliance.\n\n{% button href=\"\u003Chttps://www.dastra.eu/en/contacts/demo>\" text=\"Book a demo here\" target=\"\\_blank\" role=\"button\" class=\"btn btn-primary\" %}\n\n---\n\n## Sources\n\n- CNIL and CIANum, [*IA agentique et protection des données personnelles : équation à inconnues multiples pour les utilisateurs*](https://www.cnil.fr/sites/default/files/2026-07/ia-cianum-cnil.pdf) (20 July 2026)\n- ANSSI, [*Securing an agentic AI system*](https://messervices.cyber.gouv.fr/guides/essentiels-ia-agentique) (30 October 2026)\n- AEPD, [*Inteligencia artificial agéntica desde la perspectiva de protección de datos*](https://www.aepd.es/prensa-y-comunicacion/notas-de-prensa/la-agencia-publica-unas-orientaciones-sobre-inteligencia) (February 2026)\n- ICO, [*Tech Futures: Agentic AI*](https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/ico-tech-futures-agentic-ai) (14 January 2026), and [*Agentic AI call for evidence*](https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence/) (October 2026)\n- EDPS, [*TechDispatch 2/2025: Human oversight of automated decision-making*](https://www.edps.europa.eu/data-protection/our-work/publications/techdispatch/2025-09-23-techdispatch-22025-human-oversight-automated-making_en) (September 2025)\n- PCPD Hong Kong, [*Model Personal Data Protection Framework, Protecting Personal Data Privacy in the Use of Agentic AI*](https://www.pcpd.org.hk/english/resources_centre/publications/files/pcpd_use_of_agentic_ai.pdf) (August 2026)\n- IMDA Singapore, [*Model AI Governance Framework for Agentic AI*](https://www.imda.gov.sg/assets/d1e45a56-d898-4b9f-972e-8ae94be1a5a5.pdf) (version 1.5, May 2026, updated June 2026)\n- NIST, [*Artificial Intelligence Risk Management Framework (AI RMF 1.0)*](https://www.nist.gov/itl/ai-risk-management-framework) (January 2023), and *Generative AI Profile (NIST AI 600-1)* (July 2024)\n- ISO/IEC, [*ISO/IEC 42001:2023, Artificial intelligence – Management system*](https://www.iso.org/standard/42001) (2023)\n- European Commission, [*Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act*](https://eur-lex.europa.eu/eli/C/2026/6005/oj) (OJ C/2026/6005, 12 November 2026)","\u003Cp>AI agents no longer just answer questions. They plan, call tools, keep a memory and act: they send emails, update records, screen candidates or trigger payments, sometimes before anyone reviews the result.\u003C/p>\n\u003Cp>For privacy and compliance teams, this changes the question. It is no longer only who can use AI, but what an agent is allowed to reach, what it is allowed to do, and who answers for the result.\u003C/p>\n\u003Cp>To this date, no law has been written specifically for AI agents. However, data protection authorities, cybersecurity agencies and public bodies in Europe, the UK and Asia have published their guidance on agentic AI, and international standards such as the NIST AI RMF and ISO/IEC 42001 provide a common method. Each covers part of the picture. Read together, they show a clear consensus on what organisations will be expected to do.\u003C/p>\n\u003Cp>This article maps those frameworks: who published what, what each one focuses on, and where they agree.\u003C/p>\n\u003Ch2 id=\"why-ai-agents-need-their-own-governance\">Why AI agents need their own governance\u003C/h2>\n\u003Cp>Agents raise the same legal questions as other AI systems, but they raise the stakes in three ways:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Autonomy.\u003C/strong> An agent decides how to reach a goal, so its actions can go beyond what the user intended.\u003C/li>\n\u003Cli>\u003Cstrong>Access.\u003C/strong> To be useful, agents often need broad rights over browsers, mailboxes, files, calendars and business tools.\u003C/li>\n\u003Cli>\u003Cstrong>Action.\u003C/strong> Agents have real-world effects, often before a human sees the file.\u003C/li>\n\u003C/ul>\n\u003Ch2 id=\"the-risks-regulators-agree-on\">The risks regulators agree on\u003C/h2>\n\u003Cp>Regulators use different words, but they describe the same set of risks.\u003C/p>\n\u003Cp>\u003Cstrong>1. Roles that shift within a single task.\u003C/strong> In a classic vendor setup, it is fairly clear who decides and who acts on instructions. With agents, roles can change from one step to the next: the agent calls a third-party API, queries an external database or chains a sub-agent. A supplier described as a processor in the contract can become an independent or joint controller for part of the task. The more actors involved, the harder it is to say who answers for what.\u003C/p>\n\u003Cp>\u003Cstrong>2. Purpose drift.\u003C/strong> Purpose limitation is tied to lawful basis: an agent that goes beyond its stated purpose is likely outside its lawful basis too. An agent asked to screen CVs for technical skills may also infer age from graduation dates or \"culture fit\" from social media. That creates new purposes, with no legal basis and no information to the candidate. The problem can also start upstream, when purposes are defined too broadly for open-ended, general-purpose agents.\u003C/p>\n\u003Cp>\u003Cstrong>3. Excessive access.\u003C/strong> Agents often get broad default rights over files, emails, credentials and third-party accounts. Every source the agent can reach is personal data it may process, store or pass on. Excessive permissions also make the agent a target: anyone who compromises it inherits its access. And an action can be authorised and still cause harm, because an agent can make mistakes or combine permitted operations in a dangerous way.\u003C/p>\n\u003Cp>\u003Cstrong>4. Inferences and persistent memory.\u003C/strong> Agents draw conclusions nobody asked for. Some of them fall into sensitive categories, such as health or religion, without any legal basis covering them. Agents also remember: their persistent memory can build a detailed profile of a person, spread across several agents. Memories, histories and logs become stores of personal data in their own right, each needing a retention period.\u003C/p>\n\u003Cp>\u003Cstrong>5. Errors that spread.\u003C/strong> A generative tool that makes a mistake produces a wrong answer. An agent that makes a mistake can act on it, pass it to another agent and build on it. Inaccurate or hallucinated data can spread across a multi-agent system without alerting anyone, and the same request may not produce the same actions twice.\u003C/p>\n\u003Cp>\u003Cstrong>6. Oversight that is not effective.\u003C/strong> A human reviewer does not make a system safe by itself. Reviewers trust outputs that have been reliable in the past, approve too quickly under time pressure, and lack the information or authority to say no. A click-through approval can leave a decision effectively automated.\u003C/p>\n\u003Cp>\u003Cstrong>7. A wider attack surface.\u003C/strong> Agents read untrusted content, such as emails, documents and web pages, and can act on it. A prompt injection, a malicious instruction hidden in that content, can lead to real actions: data sent to an external server, records changed, messages sent. Plugins and connectors that have not been security-reviewed add further risk.\u003C/p>\n\u003Cp>\u003Cstrong>8. Harder rights and transparency.\u003C/strong> The more agents, memories and third-party services involved, the harder it is to tell people how their data is used. It is also harder to answer an access or erasure request completely. Which agent collected the data, where it is stored, and which service received it are questions many organisations cannot yet answer.\u003C/p>\n\u003Ch2 id=\"the-legal-baseline-the-eu-example\">The legal baseline: the EU example\u003C/h2>\n\u003Cp>Before looking at the guidance, it is worth remembering that the law already applies. Unlike the frameworks below, it is not voluntary.\u003C/p>\n\u003Cp>\u003Cstrong>Data protection law applies in full.\u003C/strong> No jurisdiction has created a special regime for agents yet. However, agents are not legal persons. The organisation that decides why and how data is processed remains accountable, whether or not an agent does the processing. In Europe for example, this means the GDPR, including its rules on automated decision-making (Article 22).\u003C/p>\n\u003Cp>\u003Cstrong>The EU AI Act applies as law.\u003C/strong> The AI Act does not define AI agents, but agents fall under its existing rules depending on how they are used:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Transparency (Article 50)\u003C/strong> has applied since 2 August 2026 to agents that interact with people or generate content.\u003C/li>\n\u003Cli>\u003Cstrong>High-risk obligations\u003C/strong> apply where an agent falls into a high-risk category, for example in recruitment or credit.\u003C/li>\n\u003Cli>\u003Cstrong>Prohibited practices (Article 5)\u003C/strong>, such as harmful manipulation or the exploitation of vulnerabilities, apply to agents like any other AI system.\u003C/li>\n\u003Cli>\u003Cstrong>AI literacy (Article 4)\u003C/strong> requires organisations to train the staff who use agents.\u003C/li>\n\u003C/ul>\n\u003Ch2 id=\"international-standards-the-method-and-the-structure\">International standards: the method and the structure\u003C/h2>\n\u003Cp>Two international references are not specific to agents, but they give organisations a common way to govern them alongside the law.\u003C/p>\n\u003Cp>\u003Cstrong>The NIST AI Risk Management Framework (AI RMF)\u003C/strong> Published by the US National Institute of Standards and Technology in January 2023, the AI RMF is voluntary guidance. It organises AI risk management around four functions:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Govern:\u003C/strong> set roles, policies and accountability;\u003C/li>\n\u003Cli>\u003Cstrong>Map:\u003C/strong> understand the context and the risks of each use case;\u003C/li>\n\u003Cli>\u003Cstrong>Measure:\u003C/strong> assess and test those risks;\u003C/li>\n\u003Cli>\u003Cstrong>Manage:\u003C/strong> treat them and monitor them over time.\u003C/li>\n\u003C/ul>\n\u003Cp>In July 2024, NIST added a profile for generative AI (NIST AI 600-1). In 2026, its Center for AI Standards and Innovation (CAISI) launched an AI Agent Standards Initiative to develop standards specific to agents.\u003C/p>\n\u003Cp>\u003Cstrong>ISO/IEC 42001\u003C/strong> Published in 2023 by ISO and IEC, ISO/IEC 42001 is the first certifiable standard for an AI management system. It works like ISO 27001 does for information security: policies, roles, risk assessment, controls, internal audits and continuous improvement. For agents, it gives a structure to keep the inventory, risk assessments and controls up to date as agents change.\u003C/p>\n\u003Cp>\u003Cstrong>How they fit together\u003C/strong>\u003Cbr />\nIn practice, an organisation can use the NIST AI RMF to assess the risks of each agent, ISO/IEC 42001 to organise its governance, and the AI Act and data protection law as the requirements that must be met.\u003C/p>\n\u003Ch2 id=\"the-regulatory-frameworks-at-a-glance\">The regulatory frameworks at a glance\u003C/h2>\n\u003Cp>The frameworks below do not create new obligations. They show how regulators expect these rules to be applied in practice.\u003C/p>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Framework\u003C/th>\n\u003Cth>Publisher\u003C/th>\n\u003Cth>Date\u003C/th>\n\u003Cth>Main focus\u003C/th>\n\u003C/tr>\n\u003C/thead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>Exploratory note on agentic AI\u003C/td>\n\u003Ctd>CNIL and CIANum (France)\u003C/td>\n\u003Ctd>July 2026\u003C/td>\n\u003Ctd>Data protection risks, loss of user control\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Cem>Securing an agentic AI system\u003C/em>\u003C/td>\n\u003Ctd>ANSSI (France)\u003C/td>\n\u003Ctd>October 2026\u003C/td>\n\u003Ctd>Security principles for deploying agents\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Agentic AI from a data protection perspective\u003C/td>\n\u003Ctd>AEPD (Spain)\u003C/td>\n\u003Ctd>February 2026\u003C/td>\n\u003Ctd>Detailed technical and GDPR analysis\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Tech Futures: Agentic AI, and call for evidence\u003C/td>\n\u003Ctd>ICO (United Kingdom)\u003C/td>\n\u003Ctd>January and October 2026\u003C/td>\n\u003Ctd>Novel data protection risks\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>TechDispatch on human oversight\u003C/td>\n\u003Ctd>EDPS (EU)\u003C/td>\n\u003Ctd>September 2025\u003C/td>\n\u003Ctd>Limits of human-in-the-loop\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Model Personal Data Protection Framework for agentic AI\u003C/td>\n\u003Ctd>PCPD (Hong Kong)\u003C/td>\n\u003Ctd>August 2026\u003C/td>\n\u003Ctd>Data protection principles and security checklist\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Model AI Governance Framework for Agentic AI\u003C/td>\n\u003Ctd>IMDA (Singapore)\u003C/td>\n\u003Ctd>v1.5, May 2026\u003C/td>\n\u003Ctd>Operational governance, with case studies\u003C/td>\n\u003C/tr>\n\u003C/tbody>\n\u003C/table>\n\u003Ch2 id=\"european-union\">European Union\u003C/h2>\n\u003Ch3 id=\"the-edps-a-human-in-the-loop-is-not-enough-on-its-own\">The EDPS: a human in the loop is not enough on its own\u003C/h3>\n\u003Cp>The European Data Protection Supervisor's TechDispatch on human oversight of automated decision-making challenges a common assumption: adding a human reviewer does not make a system safe by itself. The EDPS identifies flawed assumptions behind many oversight designs:\u003C/p>\n\u003Cul>\n\u003Cli>that the system will recognise its own limits and hand over to a human;\u003C/li>\n\u003Cli>that the reviewer will not be influenced by the output (automation bias);\u003C/li>\n\u003Cli>that the reviewer will always have the time, expertise and authority to override it.\u003C/li>\n\u003C/ul>\n\u003Cp>Its recommendations include enough review time, training, interfaces that make intervention easy, and regular audits of how reviewers decide.\u003C/p>\n\u003Ch2 id=\"france\">France\u003C/h2>\n\u003Ch3 id=\"the-cnil-and-cianum-loss-of-control-at-the-heart-of-the-analysis\">The CNIL and CIANum: loss of control at the heart of the analysis\u003C/h3>\n\u003Cp>Published in July 2026 with the French AI and Digital Council, the CNIL's exploratory note describes an \"equation with multiple unknowns\" for users. Its main concerns:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Persistent memory\u003C/strong> that builds a deep, evolving profile of the person, spread across several agents.\u003C/li>\n\u003Cli>\u003Cstrong>Autonomous execution\u003C/strong> that may diverge from what the user intended.\u003C/li>\n\u003Cli>\u003Cstrong>Unclear liability\u003C/strong> when many actors are involved.\u003C/li>\n\u003Cli>\u003Cstrong>Data spread across agents\u003C/strong>, which makes rights to access, rectification and erasure harder to exercise.\u003C/li>\n\u003C/ul>\n\u003Cp>The note is not binding. It confirms that the GDPR applies in full, and recommends four safeguards:\u003C/p>\n\u003Cul>\n\u003Cli>detecting and filtering misuse;\u003C/li>\n\u003Cli>separating memory by agent and by processing;\u003C/li>\n\u003Cli>keeping human oversight over critical actions;\u003C/li>\n\u003Cli>a kill switch.\u003C/li>\n\u003C/ul>\n\u003Ch3 id=\"anssi-authorised-does-not-mean-safe\">ANSSI: authorised does not mean safe\u003C/h3>\n\u003Cp>France's national cybersecurity agency published \u003Cem>Securing an agentic AI system\u003C/em> on 30 October 2026. It concerns any organisation that deploys or integrates an agent, including companies that buy one off the shelf and connect it to their own tools.\u003C/p>\n\u003Cp>Its starting point: an action the agent is authorised to take can still cause harm, because an agent can make mistakes, be manipulated, or combine permitted operations in a dangerous way. Its seven principles cover:\u003C/p>\n\u003Cul>\n\u003Cli>end-to-end risk analysis;\u003C/li>\n\u003Cli>limiting autonomy;\u003C/li>\n\u003Cli>sorting actions into autonomous, subject to approval, or prohibited;\u003C/li>\n\u003Cli>separating environments;\u003C/li>\n\u003Cli>controls that do not depend on the model;\u003C/li>\n\u003Cli>limiting the impact of the system;\u003C/li>\n\u003Cli>demonstrating control over time.\u003C/li>\n\u003C/ul>\n\u003Cp>For legal teams, the security requirements from the risk analysis should appear in the specifications of any product or service purchased.\u003C/p>\n\u003Ch2 id=\"spain\">Spain\u003C/h2>\n\u003Ch3 id=\"the-aepd-a-detailed-analysis-of-risks-specific-to-agents\">The AEPD: a detailed analysis of risks specific to agents\u003C/h3>\n\u003Cp>The Spanish authority's guidance, published in February 2026, focuses on what is specific to agents beyond the risks already known from generative AI:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Excessive permissions.\u003C/strong> They turn agents into pivot points for attackers.\u003C/li>\n\u003Cli>\u003Cstrong>Errors that compound\u003C/strong> over long tasks.\u003C/li>\n\u003Cli>\u003Cstrong>Memories and logs that become large stores of personal data.\u003C/strong>\u003C/li>\n\u003Cli>\u003Cstrong>\"BYOAgentic\".\u003C/strong> Employees build their own agents outside any governance.\u003C/li>\n\u003C/ul>\n\u003Cp>The AEPD also warns that relying entirely on users to catch errors is itself a threat. It recommends deciding the level of autonomy for each processing operation, from \"the agent proposes, the human acts\" to \"the agent acts, the human observes\".\u003C/p>\n\u003Ch2 id=\"united-kingdom\">United Kingdom\u003C/h2>\n\u003Ch3 id=\"the-ico-eight-novel-data-protection-risks\">The ICO: eight novel data protection risks\u003C/h3>\n\u003Cp>In its Tech Futures report of January 2026, the ICO mapped the data protection risks specific to agentic AI:\u003C/p>\n\u003Cul>\n\u003Cli>unclear controller and processor roles across the supply chain;\u003C/li>\n\u003Cli>more automated decision-making;\u003C/li>\n\u003Cli>purposes set too broadly for open-ended agents;\u003C/li>\n\u003Cli>processing beyond what is necessary;\u003C/li>\n\u003Cli>unintended inference of special category data;\u003C/li>\n\u003Cli>transparency and rights made harder by complexity;\u003C/li>\n\u003Cli>new cyber threats;\u003C/li>\n\u003Cli>the concentration of personal data in personal assistant agents.\u003C/li>\n\u003C/ul>\n\u003Cp>In October 2026, the ICO opened a call for evidence on agentic AI, closing on 20 November 2026. It seeks evidence on several topics such as agent identities and permissions or meaningful human oversight.\u003C/p>\n\u003Cp>The responses will feed into future guidance and the ICO's statutory code of practice on AI and automated decision-making.\u003C/p>\n\u003Ch2 id=\"asia\">Asia\u003C/h2>\n\u003Ch3 id=\"the-pcpd-hong-kong-nine-recommendations-and-a-security-checklist\">The PCPD (Hong Kong): nine recommendations and a security checklist\u003C/h3>\n\u003Cp>Published in August 2026, the PCPD's framework applies Hong Kong's six Data Protection Principles to agents. Its premise is shared by every regulator: AI agents are not legal persons, and the organisation remains accountable.\u003C/p>\n\u003Cp>It identifies five risks:\u003C/p>\n\u003Cul>\n\u003Cli>extensive default access;\u003C/li>\n\u003Cli>system vulnerabilities;\u003C/li>\n\u003Cli>unvetted plugins and skills;\u003C/li>\n\u003Cli>function creep;\u003C/li>\n\u003Cli>errors cascading between agents.\u003C/li>\n\u003C/ul>\n\u003Cp>Its nine recommendations range from avoiding data collected \"just in case\" to keeping a human in the loop for high-stakes or irreversible actions. A practical security checklist covers each stage of an agent's life, from evaluation to uninstallation.\u003C/p>\n\u003Ch3 id=\"imda-singapore-the-operational-framework\">IMDA (Singapore): the operational framework\u003C/h3>\n\u003Cp>Singapore's Model AI Governance Framework for Agentic AI, now in version 1.5, is built around four dimensions:\u003C/p>\n\u003Col>\n\u003Cli>\u003Cstrong>Assess and bound the risks upfront.\u003C/strong> Score each use case on impact and likelihood, then limit what the agent can reach.\u003C/li>\n\u003Cli>\u003Cstrong>Make humans meaningfully accountable.\u003C/strong> Allocate responsibilities across the value chain and set approval checkpoints.\u003C/li>\n\u003Cli>\u003Cstrong>Implement technical controls and processes.\u003C/strong> Test whole workflows, roll out gradually and monitor continuously.\u003C/li>\n\u003Cli>\u003Cstrong>Enable end-user responsibility.\u003C/strong> Tell users what the agent can do, train staff, and protect the skills they need to spot errors.\u003C/li>\n\u003C/ol>\n\u003Cp>Two ideas from IMDA are useful well beyond Singapore. The first distinguishes an agent's action-space (what it can do) from its autonomy (how freely it decides). The second is concrete measures against automation bias, such as tracking how often reviewers override the agent and how long they take to decide. The framework is illustrated with case studies from banks, technology companies and public bodies.\u003C/p>\n\u003Ch2 id=\"where-the-frameworks-agree\">Where the frameworks agree\u003C/h2>\n\u003Cp>Across very different jurisdictions, the same expectations come back:\u003C/p>\n\u003Col>\n\u003Cli>\u003Cstrong>Accountability stays with the organisation.\u003C/strong> Agents are not legal persons, and someone must answer, by name, for what each agent does.\u003C/li>\n\u003Cli>\u003Cstrong>Know your agents.\u003C/strong> Every agent should be listed, with its owner, its purpose, its access and its risk level.\u003C/li>\n\u003Cli>\u003Cstrong>Scale controls to the risk.\u003C/strong> Autonomy, access to sensitive data and the reversibility of actions are the main factors.\u003C/li>\n\u003Cli>\u003Cstrong>Limit access to what the task needs.\u003C/strong> Enforce it in the system, not only through instructions in the prompt.\u003C/li>\n\u003Cli>\u003Cstrong>Bound memory and inferences.\u003C/strong> Set retention periods, filter sensitive inferences, and make sure rights work across every memory.\u003C/li>\n\u003Cli>\u003Cstrong>Make human oversight real.\u003C/strong> Put checkpoints on high-stakes actions, and give reviewers the time and authority to say no.\u003C/li>\n\u003Cli>\u003Cstrong>Trace and monitor.\u003C/strong> Keep logs that reconstruct each task, test before go-live, and reassess after every change.\u003C/li>\n\u003C/ol>\n\u003Ch2 id=\"what-it-means-for-privacy-and-compliance-teams\">What it means for privacy and compliance teams\u003C/h2>\n\u003Cp>These frameworks are not binding. But they show how existing rules, starting with the GDPR and the AI Act, will be read and enforced when agents are involved. Organisations that wait for agent-specific legislation are taking a risk.\u003C/p>\n\u003Cp>Three first steps to take:\u003C/p>\n\u003Col>\n\u003Cli>\u003Cstrong>Find out which agents are already running.\u003C/strong> Include features switched on in existing software and agents built by employees.\u003C/li>\n\u003Cli>\u003Cstrong>Check your existing governance.\u003C/strong> Make sure your policies, AI register, contracts and impact assessments cover what agents actually do.\u003C/li>\n\u003Cli>\u003Cstrong>Assess before deploying.\u003C/strong> Do it at the start of each project, not after go-live.\u003C/li>\n\u003C/ol>\n\u003Cp>\u003Cstrong>Go further\u003C/strong> Dastra helps privacy and AI compliance teams govern their AI agents in one place: build your AI register, assess suppliers and use cases, track risks and measures, and connect everything to your GDPR compliance.\u003C/p>\n\u003Cdiv class=\"content-btn-container\">\u003Ca href=\"\u003Ca href=\" target=\"_blank\" role=\"button\" class=\"btn btn-primary\" rel=\"nofollow\">Book a demo here\u003C/a>\u003C/div>\n\u003Chr />\n\u003Ch2 id=\"sources\">Sources\u003C/h2>\n\u003Cul>\n\u003Cli>CNIL and CIANum, \u003Ca href=\"https://www.cnil.fr/sites/default/files/2026-07/ia-cianum-cnil.pdf\" rel=\"nofollow\">\u003Cem>IA agentique et protection des données personnelles : équation à inconnues multiples pour les utilisateurs\u003C/em>\u003C/a> (20 July 2026)\u003C/li>\n\u003Cli>ANSSI, \u003Ca href=\"https://messervices.cyber.gouv.fr/guides/essentiels-ia-agentique\" rel=\"nofollow\">\u003Cem>Securing an agentic AI system\u003C/em>\u003C/a> (30 October 2026)\u003C/li>\n\u003Cli>AEPD, \u003Ca href=\"https://www.aepd.es/prensa-y-comunicacion/notas-de-prensa/la-agencia-publica-unas-orientaciones-sobre-inteligencia\" rel=\"nofollow\">\u003Cem>Inteligencia artificial agéntica desde la perspectiva de protección de datos\u003C/em>\u003C/a> (February 2026)\u003C/li>\n\u003Cli>ICO, \u003Ca href=\"https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/ico-tech-futures-agentic-ai\" rel=\"nofollow\">\u003Cem>Tech Futures: Agentic AI\u003C/em>\u003C/a> (14 January 2026), and \u003Ca href=\"https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence/\" rel=\"nofollow\">\u003Cem>Agentic AI call for evidence\u003C/em>\u003C/a> (October 2026)\u003C/li>\n\u003Cli>EDPS, \u003Ca href=\"https://www.edps.europa.eu/data-protection/our-work/publications/techdispatch/2025-09-23-techdispatch-22025-human-oversight-automated-making_en\" rel=\"nofollow\">\u003Cem>TechDispatch 2/2025: Human oversight of automated decision-making\u003C/em>\u003C/a> (September 2025)\u003C/li>\n\u003Cli>PCPD Hong Kong, \u003Ca href=\"https://www.pcpd.org.hk/english/resources_centre/publications/files/pcpd_use_of_agentic_ai.pdf\" rel=\"nofollow\">\u003Cem>Model Personal Data Protection Framework, Protecting Personal Data Privacy in the Use of Agentic AI\u003C/em>\u003C/a> (August 2026)\u003C/li>\n\u003Cli>IMDA Singapore, \u003Ca href=\"https://www.imda.gov.sg/assets/d1e45a56-d898-4b9f-972e-8ae94be1a5a5.pdf\" rel=\"nofollow\">\u003Cem>Model AI Governance Framework for Agentic AI\u003C/em>\u003C/a> (version 1.5, May 2026, updated June 2026)\u003C/li>\n\u003Cli>NIST, \u003Ca href=\"https://www.nist.gov/itl/ai-risk-management-framework\" rel=\"nofollow\">\u003Cem>Artificial Intelligence Risk Management Framework (AI RMF 1.0)\u003C/em>\u003C/a> (January 2023), and \u003Cem>Generative AI Profile (NIST AI 600-1)\u003C/em> (July 2024)\u003C/li>\n\u003Cli>ISO/IEC, \u003Ca href=\"https://www.iso.org/standard/42001\" rel=\"nofollow\">\u003Cem>ISO/IEC 42001:2023, Artificial intelligence – Management system\u003C/em>\u003C/a> (2023)\u003C/li>\n\u003Cli>European Commission, \u003Ca href=\"https://eur-lex.europa.eu/eli/C/2026/6005/oj\" rel=\"nofollow\">\u003Cem>Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act\u003C/em>\u003C/a> (OJ C/2026/6005, 12 November 2026)\u003C/li>\n\u003C/ul>\n","AI Agent Governance: Frameworks Privacy & Compliance Teams Need","How should organisations govern AI agents? An overview of regulatory frameworks from Europe, the UK and Asia, and the controls privacy teams need now.",2852,16,"Governing AI agents: the frameworks every privacy & compliance team should know",0,null,"en","ai-agent-governance-frameworks","Published",{"id":18,"displayName":19,"avatarUrl":20,"bio":13,"blogUrl":13,"color":13,"userId":18,"creationDate":21},20352,"Leïla Sayssa","https://static.dastra.eu/tenant-3/avatar/20352/TDYeY3C8Rz1lLE/dpo-avatar-h01-150.png","2025-03-03T11:08:22","2026-11-13T13:16:00","2026-11-13T13:16:22.1079067","2026-11-19T12:42:52.2224199",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":31},2,"Blog","A list of curated articles provided by the community","blog","#28449a",[32,35,38],{"lang":33,"name":27,"description":34},"fr","Une liste d'articles rédigés par la communauté",{"lang":36,"name":27,"description":37},"es","Una lista de artículos escritos por la comunidad",{"lang":39,"name":27,"description":40},"de","Eine Liste von Artikeln, die von der Community verfasst wurden",[42,47],{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":43},[44,45,46],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},{"id":48,"name":49,"description":50,"url":51,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":53,"order":58,"translations":59},221,"AI Governance","Best practices, regulatory frameworks and real-world insights to manage AI responsibly and in compliance with the AI Act.","ai-governance","#000000",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":54},[55,56,57],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},3,[60,63],{"lang":33,"name":61,"description":62},"Gouvernance de l'IA","Bonnes pratiques, cadres réglementaires et retours d'expérience pour piloter l'IA de façon responsable et conforme à l'AI Act.",{"lang":39,"name":64,"description":65},"KI-Governance","Best Practices, regulatorische Rahmenbedingungen und Praxiserfahrungen für einen verantwortungsvollen KI-Einsatz im Einklang mit dem AI Act.",[],"https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-original.png",[69,70,71,72,73,74,75],"https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-1000.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-1500.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-800.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-600.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-300.webp","https://static.dastra.eu/content/74830204-9f9a-4bde-98cf-1861970b5bf0/ai-agents-100.webp",61724,{"total":78,"items":79,"parent":228},8,[80,101,121,143,165,186,195,211],{"id":81,"name":82,"description":83,"url":84,"color":85,"parentId":26,"count":13,"imageUrl":13,"parent":86,"order":78,"translations":91},20,"Inside Dastra","Go behind the scenes at Dastra: company news, culture, events, team highlights, and the people driving our GDPR solution.","dastra-life","#e3cf68",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":87},[88,89,90],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},[92,95,98],{"lang":33,"name":93,"description":94},"Vie de Dastra","Plongez dans les coulisses de Dastra : actualités internes, culture d’entreprise, événements, équipes et engagements. Découvrez qui se cache derrière notre solution RGPD.",{"lang":39,"name":96,"description":97},"Innerhalb von Dastra","Eintauchen in das Unternehmen",{"lang":36,"name":99,"description":100},"Dentro de Dastra","Sumérjase en la empresa",{"id":102,"name":103,"description":104,"url":105,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":106,"order":111,"translations":112},69,"Expertise","Gain insights from our experts on GDPR compliance, data protection, and privacy challenges. In-depth articles, professional analysis, and real-world best practices.","indepth",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":107},[108,109,110],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},7,[113,115,118],{"lang":33,"name":103,"description":114},"Bénéficiez des conseils de nos experts sur la conformité RGPD, la protection des données et les enjeux privacy. Articles de fond, analyses et retours d’expérience métier.",{"lang":39,"name":116,"description":117},"Fachwissen","Entdecken Sie die Artikel unserer DSGVO-Experten",{"lang":36,"name":119,"description":120},"Experiencia","Descubre los artículos de nuestros expertos en Privacy",{"id":122,"name":123,"description":124,"url":125,"color":126,"parentId":26,"count":13,"imageUrl":13,"parent":127,"order":132,"translations":133},4,"Use cases","Discover how companies use Dastra to manage their GDPR compliance. Testimonials, use cases, and real-world integrations of our privacy management solution.","case-studies","#b61b9c",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":128},[129,130,131],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},6,[134,137,140],{"lang":33,"name":135,"description":136},"Retours d'expérience","Découvrez comment les entreprises utilisent Dastra pour piloter leur conformité RGPD. Témoignages, cas d’usage et intégrations concrètes de notre solution de privacy management.",{"lang":39,"name":138,"description":139},"Case Studies","Entdecken Sie die Erfahrungsberichte unserer Kunden",{"lang":36,"name":141,"description":142},"Casos prácticos","Descubra los testimonios de nuestros clientes",{"id":144,"name":145,"description":146,"url":147,"color":148,"parentId":26,"count":13,"imageUrl":13,"parent":149,"order":154,"translations":155},10,"Release notes","Keep up with the latest features of Dastra: product updates, new functionalities, and improvements to our GDPR compliance and data management platform.","release","#8c316a",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":150},[151,152,153],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},5,[156,159,162],{"lang":33,"name":157,"description":158},"Notes de version","Restez informé(e) des dernières fonctionnalités de Dastra : mises à jour, évolutions produit et améliorations de notre solution de conformité RGPD et de gestion des données personnelles.",{"lang":39,"name":160,"description":161},"Veröffentlichungen","Alle Versionshinweise und Funktionen von Dastra",{"lang":36,"name":163,"description":164},"Lanzamientos","Todas las notas de la versión e información sobre las funciones de Dastra",{"id":166,"name":167,"description":168,"url":169,"color":170,"parentId":26,"count":13,"imageUrl":13,"parent":171,"order":122,"translations":176},9,"News","Stay up to date with the latest news from data protection authorities: decisions, fines, guidelines, and regulatory trends in GDPR and privacy.","news","#1676ca",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":172},[173,174,175],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},[177,180,183],{"lang":33,"name":178,"description":179},"Actualités","Suivez les dernières actualités des autorités de protection des données (CNIL, EDPS, etc.) : décisions, sanctions, lignes directrices et tendances réglementaires en matière de RGPD et de privacy.",{"lang":36,"name":181,"description":182},"Actualidad","Todos los artículos relativos a las autoridades de protección de datos",{"lang":39,"name":184,"description":185},"Nachrichten","Alle Artikel mit Bezug zu Datenschutzbehörden",{"id":48,"name":49,"description":50,"url":51,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":187,"order":58,"translations":192},{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":188},[189,190,191],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},[193,194],{"lang":33,"name":61,"description":62},{"lang":39,"name":64,"description":65},{"id":196,"name":197,"description":198,"url":199,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":200,"order":26,"translations":205},220,"Compliance","Regulatory news, practical guides and analysis to keep your organization compliant with the GDPR and beyond.","compliance",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":201},[202,203,204],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},[206,209],{"lang":33,"name":207,"description":208},"Conformité","Actualités réglementaires, guides pratiques et analyses pour maintenir votre organisation en conformité RGPD et au-delà.",{"lang":39,"name":197,"description":210},"Regulatorische Neuigkeiten, Praxisleitfäden und Analysen, um Ihre Organisation DSGVO-konform und darüber hinaus aufzustellen.",{"id":212,"name":213,"description":214,"url":215,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":216,"order":221,"translations":222},219,"Privacy","Principles, techniques and trends in personal data protection — from privacy by design to data subject rights.","privacy",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":217},[218,219,220],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},1,[223,225],{"lang":33,"name":213,"description":224},"Principes, techniques et tendances autour de la protection des données personnelles — de la privacy by design aux droits des personnes.",{"lang":39,"name":226,"description":227},"Datenschutz","Grundsätze, Methoden und Trends zum Schutz personenbezogener Daten — von Privacy by Design bis zu Betroffenenrechten.",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":229},[230,231,232],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},[],{"items":235,"total":305,"size":221,"page":221},[236],{"title":237,"nbDownloads":12,"excerpt":13,"lang":14,"url":238,"intro":239,"featured":4,"state":16,"author":240,"authorId":18,"datePublication":241,"dateCreation":242,"dateUpdate":243,"mainCategory":244,"categories":250,"metaDatas":271,"imageUrl":295,"imageThumbUrls":296,"id":304},"AI Act: Is Your Organization Ready for December 2, 2026?","ai-act-checklist-en","Are you ready for 2 December 2026? Article 50(2), new prohibitions, and 6 checks to carry out. Download the free AI Act checklist.",{"id":18,"displayName":19,"avatarUrl":20,"bio":13,"blogUrl":13,"color":13,"userId":18,"creationDate":21},"2026-11-17T15:07:00","2026-11-17T15:07:01.2688405","2026-11-19T12:41:21.9944427",{"id":245,"name":246,"description":13,"url":247,"color":248,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":58,"translations":249},70,"Livre blanc","white-papers","#1795d3",[],[251,260,265,267],{"id":221,"name":252,"description":253,"url":254,"color":255,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":221,"translations":256},"News and watch","All the news about the GDPR compliance in application","watch","#6c1414",[257],{"lang":33,"name":258,"description":259},"Actualités & veille","Toutes les dernières actualités sur la conformité GDPR dans l'application.",{"id":26,"name":27,"description":28,"url":29,"color":30,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":12,"translations":261},[262,263,264],{"lang":33,"name":27,"description":34},{"lang":36,"name":27,"description":37},{"lang":39,"name":27,"description":40},{"id":245,"name":246,"description":13,"url":247,"color":248,"parentId":13,"count":13,"imageUrl":13,"parent":13,"order":58,"translations":266},[],{"id":48,"name":49,"description":50,"url":51,"color":52,"parentId":26,"count":13,"imageUrl":13,"parent":13,"order":58,"translations":268},[269,270],{"lang":33,"name":61,"description":62},{"lang":39,"name":64,"description":65},[272,275],{"typeMetaDataId":122,"value":273,"id":274},"https://static.dastra.eu/backofficefilescontainer/bc880b52-7ec0-4e3d-88cc-ab153c4c0e40/Checklist AI Act 2 December - Dastra EN.pdf",120483,{"typeMetaDataId":166,"value":276,"id":294},{"title":277,"items":278},"Frequently asked questions about the AI Act and the Dec. 2, 2026 deadline",[279,282,285,288,291],{"question":280,"answer":281},"Which obligations apply on 2 December 2026?","\u003Cp>The machine-readable marking of synthetic content (Art. 50(2)) for systems placed on the market before 2 August 2026, and two new prohibitions under Article 5.\u003C/p>\u003Cp>\u003C/p>",{"question":283,"answer":284},"Is a deployer concerned by Article 50(2)?","\u003Cp>No, this obligation falls on providers. Deployers have other transparency obligations under Article 50, applicable since 2 August 2026, and remain subject to the Article 5 prohibitions.\u003C/p>\u003Cp>\u003C/p>",{"question":286,"answer":287},"My system was launched after 2 August 2026: do I get a grace period?","\u003Cp>No. The 2 December 2026 deadline only covers systems already on the market before 2 August 2026.\u003C/p>\u003Cp>\u003C/p>",{"question":289,"answer":290},"What are the penalties for non-compliance?","\u003Cp>An infringement of Article 5 can reach €35 million or 7% of worldwide turnover. For Article 50, the cap is €15 million or 3%.\u003C/p>\u003Cp>\u003C/p>",{"question":292,"answer":293},"Do the high-risk system obligations also apply?","\u003Cp>No. They apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.\u003C/p>",120484,"https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-original.png",[297,298,299,300,301,302,303],"https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-1000.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-1500.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-800.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-600.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-300.webp","https://static.dastra.eu/content/bcec9acf-a4e2-4f7b-9be2-217848ee54a1/ai-act-compliance-100.webp",61765,15]