Legal development

Data Bytes 70: Your EMEA Data Privacy Update for November 2026

Insight Hero Image

    Welcome to the November edition of Data Bytes. 

    If one theme has defined the past month, it is agentic AI. Regulators on both sides of the Channel are no longer looking only at how AI models are built. They are now asking how AI agents, which can plan, act and make decisions with growing autonomy, fit within existing data protection, cyber and AI frameworks. 

    Our Spotlight this month looks at the ICO's call for evidence on agentic AI, published on 8 October 2026. It sets out the ICO's "challenges" and "expectations" across six themes: data security, transparency, accountability, ADM, fairness and purpose limitation, and lawfulness of processing. The call for evidence closes on 20 November and will feed into the ICO's statutory code of practice on AI and ADM. In the absence of standalone UK AI legislation, that code is set to become a key pillar of AI regulation in the UK. We also consider the wider cross-regulatory picture. As the DRCF's Chief Executive put it, "AI does not arrive in neat regulatory boxes". Organisations should therefore break down their own internal silos when assessing agentic AI. 

    We also share our key takeaways from the ICO's Data Protection Practitioners' Conference, including practical guidance on handling AI-generated DSARs. In the UK In Case You Missed It section, recent ICO reprimands against the Metropolitan Police Service and Data Cymru are a timely reminder of the basics. These include the risks of "shadow IT", dormant test accounts and weak passwords (yes, "Password1"), and the importance of compliant sub-processing contracts. We also flag the Supreme Court hearing in Farley v Equiniti on whether a seriousness threshold applies to claims for non-material damage. Judgment is awaited. 

    In the EU, the Irish DPC has published insights from supervising 180 AI products and services. The AI Office has also begun enforcing the AI Act's rules for general-purpose AI models, following reported loss of control incidents. Elsewhere, the Council has agreed its negotiating position on the Digital Omnibus.  

    The common thread this month is that organisations cannot wait for complete regulatory clarity.  

    As always, please get in touch with any member of the Data Bytes team if you would like to discuss any of the issues raised. 

    Spotlight on Agentic AI and data protection: the ICO reveals its current expectations 

    The ICO’s call for evidence

    On 8 October 2026, the Information Commission’s Office (ICO) published a call for evidence on agentic AI. Its aim is to identify where the ICO can provide greater clarity and practical support in forthcoming guidance, including its statutory code of practice on AI and automated decision-making (ADM).

    The code once finalised will merit particular attention as in the absence of any standalone UK AI legislation comparable to the EU AI Act, it will ultimately form a key pillar of AI regulation in the UK.

    The call for evidence is seeking feedback from organisations before 20 November and is noteworthy as it signals the ICO’s current thinking on the topic of agentic AI. Whilst the six themes (data security, transparency, accountability, ADM, fairness and purpose limitation, and lawfulness of processing) are predictable, organisations should pay particular attention to the framing of each theme, the ICO sets out “the challenge” as well as its “expectations”. Examples of the challenges include:

    • Accountability: The ICO notes that agentic systems can involve multiple organisations, models, agents and tools, and those involved may change with the workflow, making it hard to identify who determines the purposes and means of each processing activity.
    • ADM: organisations must assess whether Articles 22A–22D of the UK GDPR apply across the complete workflow, not to agents or decision points in isolation. Human involvement at one stage does not necessarily take a decision outside the ADM provisions; the question is whether it is meaningful.
    • Security and transparency: agent autonomy create challenges with security and transparency principles and requires consideration of least-privilege access, approval thresholds, tracing and audit logging.

    The ICO is clearly attempting to demonstrate how it is attempting to think practically and realistically about how data protection law applies to AI agents. Comparable guidance is lacking at an EU level. In the absence of any European Data Protection Board opinion or guidance, developers and deployers are left to rely on piecemeal guidance from individual regulators, such as the Spanish data protection authority’s (AEPD) guide on agentic AI.

    A cross-regulatory issue

    The ICO’s call for evidence has been released against the backdrop of a wider regulatory focus on agentic AI in the UK. This is clear from the priorities and recent publications by the Digital Regulation Cooperation Forum (DRCF), which brings together the ICO, CMA, Ofcom and FCA. In a keynote speech on 23 September 2026, DRCF Chief Executive Kate Jones gave the example of an AI agent booking a consumer’s holiday which, in doing so, processes personal data, purchases financial services, takes payment and affects competition between providers. She observed that “this one agentic example has engaged the remits of all four members”, explaining that the “technology converges, whereas regulation historically specialised. AI does not arrive in neat regulatory boxes.” She also emphasised that “AI autonomy does not remove organisational responsibility”, urging organisations to “end the internal silos just as we regulators have done via the DRCF”.

    Key takeaways

    Taken together, these developments point to a common theme: the decisions and capabilities of AI agents are transcending regulatory boundaries and remits and the UK regulators are responding collaboratively, as the DRCF’s work shows. This points towards the need for organisations to start evolving the compliance assessments for agentic AI to ensure they consider legal issues holistically in accordance with regulatory expectations.

    The challenge for organisations is that may not be able to wait for absolute clarity on regulatory expectations. AI developers and deployers are making compliance decisions now, and the ICO’s call for evidence, with its stated “expectations” for each theme, alongside the ICO’s recent report on generative AI (see UK key update below) are some of the few sources currently available to guide those decisions.

    UK Key Updates

    Agentic AI: ICO secures commitments from foundation model developers and turns its attention to AI agents 

    On 8 October 2026, the ICO published a report titled Building trust and transparency into generative AI development: our work to create regulatory certainty (the Report). The Report follows a supervision programme, set up in 2025 under the ICO's AI & Biometrics Strategy, which originally targeted 11 priority foundation model developers selected by reference to compliance risk, UK market share and reliance on higher-risk training datasets.

    The Report focuses on the following data protection challenges related to GenAI model development for foundational model developers and provides clarification on ICO expectations (set out here):

    • transparency – in particular, the challenge with how you inform third parties that their data is used to train AI and as the sources and modalities of training data (e.g. image, sound, videos) continue to increase;
    • lawful basis – broad interests such as “training models” is unlikely to be a valid and specific interest for processing;
    • data subject rights – the ICO is clear that: (i) "effective mechanisms for exercising rights are likely to require a combination of dataset and model-focused measures, as well as output-focused safeguards, that work together across the AI development lifecycle"; and (ii) reliance on article 11 UK GDPR should not be applied broadly;
    • do AI models contain personal data – the ICO clarifies expectations such that the extent to which an AI model memorises training data and will therefore contain personal data requires a case by case assessment of the model and a number of factors are provided to help with this assessment e.g. size of model and training dataset and amount of duplication in training data; and
    • special category data – the ICO considers that the ‘manifestly made public’ and ‘scientific research’ processing conditions are potentially applicable to special category data processed for foundation model training.– and clarified certain expectations with guidelines on each of these topics set out here.

    The ICO provides examples of how it has seen foundation model developers implement data protection improvements (listed here) to address the above challenges and these include:

    • completing computability assessments;
    • publishing sources used in training;
    • including details about implemented safeguards in LIAs; and
    • adoption of measures such as de-duplication, filtering, testing across different languages and use of classifiers and hash matching to reduce the risks associated with use of special category data.

    The Report sets out the ICO's next steps including: 

    • developing a statutory code of practice on AI and automated decision-making (ADM);
    • releasing guidance on emerging technologies such as agentic AI;
    • continued scrutiny of foundation model developers through ongoing engagement and continued reliance on legitimate interests; and
    • engagement with government to discuss special category data use. 

    Organisations piloting or deploying AI agents should review the Report, alongside the ICO’s call for evidence as together they provide insights into the ICO’s expectations and can help steer what appropriate safeguards and risk mitigations should be in place. 

    ICO Data Protection Practitioners' Conference 2026: key takeaways 

    On 13 October 2026, the ICO held its annual Data Protection Practitioners' Conference (DPPC) online. The afternoon programme comprised twelve practical seminars. Sessions of particular relevance to in-house teams included:

    • DUAA complaints requirement: a dedicated session on the new obligation for controllers to handle data protection complaints;
    • AI and ADM: sessions on the updated ADM rules and the ICO's draft guidance, alongside an update on agentic AI guidance;
    • International transfers: a session on using the Article 49 derogations as part of the UK GDPR transfers toolkit; and
    • Cyber resilience: a panel on the human impact of cyber incidents and building organisational resilience, consistent with the ICO's recent decision to join the National Cyber Resilience Centre Group's National Ambassador programme.

    Of particular interest was the session on data subject access requests (DSARs) which focused on AI generated DSARs. Key takeaways were as follows:

    • the use of AI to create DSARs doesn't change if and how you comply with and fulfil DSARs;
    • consider what the intention of the request is and if you need clarification – ask for it;
    • typically AI generated DSARs request "all information". The ICO reminded organisations that: even if an individual explicitly asks for: (i) a particular type of data, this doesn't mean that it needs to be provided; and (ii) certain search terms to be used, this doesn't automatically need to be complied with. As organisations are only required to conduct a reasonable and proportionate search; DSARs only entitle individuals to their personal data and third-party data may be redacted (in applying the third-party balancing test) and exemptions apply, this means that individuals are highly unlikely to receive all information they request. Therefore, we recommend referring to these factors to manage expectations and to set the scene for how you're approaching DSARs in compliance with ICO guidance; and
    • document steps taken and justifications for why certain documents weren't provided, so that you have a paper trail and can explain decisions.
    • Since the date of the conference, the ICO has published its guidance on AI-generated DSARs. 

    Session recordings are available on the ICO's DPPC webpage.  

    UK | In case you missed it

    1. On 12 October 2026, AI Minister Kanishka Narayan delivered a speech at Imperial College London's Base One setting out the Government's plans to use frontier AI to strengthen cyber defences, particularly in anticipation of incidents using autonomous AI systems. Considerations of how/should the Cyber Security and Resilience Bill be strengthened to address emerging AI risks are underway. See more details here.
    2. On 20 October 2026, the NCSC published the rewritten version of its research problem book, to include a new chapter titled 'AI in cyber security problems'. The NCSC explains that as a result of the acceleration and use of AI, which can plan, reason, use tools and act with growing autonomy, this impacts the pace and scale of cyber threats and therefore many established practices (such as vulnerability management, incident response and risk management) predate this technology and may need to be evolved and updated. There are three main themes to the chapter: AI for autonomous cyber defence, Security of AI systems, and AI for cyber resilience at national scale. See more details here.
    3. On 23 September 2026, the ICO issued a reprimand and enforcement notice to the Metropolitan Police Service (MPS) for unlawful processing, lack of appropriate security and approval controls related to use of third-party technology, failure to complete a DPIA. MPS officers had used a third-party retail crime intelligence-sharing platform that had not been formally approved, with around 500 accounts, for policing activities. The case is a useful reminder for all organisations of the risks of "shadow IT" adopted by staff outside formal procurement and governance channels. See more details here.
    4. On 14 October 2026, the ICO issued reprimands to the Local Government Data Unit in Wales (Data Cymru) and its IT provider following a ransomware attack. The threat actor exploited a test account set up by the IT provider over 12 years prior with the password "Password1". In addition to breach of security obligations, Data Cymru was found to be in breach of Article 28 obligations because no compliant sub-processing contract was in place with the IT provider. The lessons from this incident include removing dormant and test accounts, blocking weak passwords, protecting remote access with MFA and clearly allocating security responsibilities in outsourcing contracts. See more details here.
    5. On 6 October 2026, ASOS confirmed that an unauthorised notification had been sent to customers through its app, addressed to the company's DPO and IT team claiming that its Snowflake environment had been compromised. This form of communication by the threat actor via customers is a novel approach in the UK and is demonstrative of the extent to which threat actors are trying to exert control and force companies to take action.
    6. On 7 and 8 October 2026, the UK Supreme Court heard the appeal in Michael Farley and 431 others v Paymaster (1836) Limited (trading as Equiniti), which concerns whether a threshold of seriousness applies to claims for compensation in relation to non-material damage under the UK GDPR. Judgment is awaited. This case concerns a 2019 data breach in which Sussex Police officers' pension statements were posted to outdated addresses and the appeal has arisen following the Court of Appeal's decision in 2025 that that the claims could not be struck out for failing to meet a threshold of seriousness. See more details here.
    7. On 9 October 2026, the ICO reported that complaints about automated nuisance calls had nearly tripled in three months and were the highest in volume since 2021. The rise of complaints is reflective of increasing concern about automated and AI-enabled “robo" calls and this report demonstrates that it often leads to enforcement action and regulatory scrutiny. Given the increased fining regime under PECR (aligning with UK GDPR fines) for marketing infringements including automated calls, this is a caution to organisations using automated calls. See more details here.

    EU Key updates

    1. Irish DPC publishes lessons from supervising AI services

    On 24 September 2026, the Irish Data Protection Commission (DPC) published its Responsible Artificial Intelligence Innovation: Insights from the Data Protection Commission's Supervision of AI (2021-2025) (DPC Report). 

    As the European lead supervisory authority for many large technology companies, the DPC Report provides invaluable insights following the DPC's engagement with and supervision of 180 AI products and services provided by a range of companies including Big Tech and covers topics including LLMs, recommender systems, age assurance, facial recognition, personalisation and agents.

    According to the DPC Report, from 2021 to 2025, the DPC Supervision Function has seen a tenfold increase in AI-related engagements with large technology contractors reflecting the increase of use and development of AI. 

    Key takeaways from the DPC Report: 

    • DPIA: Even though AI may not be "new", the creation of a new AI or AI model may still be considered a new technology that may have unknown risks that will trigger the need for a DPIA and where there is an appointed DPO, they must be involved in the creation and documentation of DPIAs.
    • Transparency: additional transparency notices should be published prior to the creation of training and testing datasets; transparency must precede processing - "no amount of post-processing transparency will relieve a controller of its obligations to provide transparency prior to processing" and the DPC notes that transparency should be given prior to training in light of the challenge that arises when trying to remove personal data from a model; controllers deploying AI are responsible for educating individuals on the data protection implications and risks; and lack of transparency can undermine the validity of a lawful basis.
    • Legitimate interests – reliance on legitimate interests for AI training is a particular focus and whilst it is acknowledged that it can be relied on, legitimate interest assessments should, amongst other things, address: whether sufficient transparency has been provided to meet data subjects' reasonable expectations – this involves identifying and understanding what their reasonable expectations would be in light of the circumstances at the time of data collection and how data has been collected; and how much friction there is for data subjects trying to exercise their rights.
    • Right to object – Acknowledging that personal data used to train a model can be very difficult to be removed at a later stage, the opportunity to opt out must be given to individuals before the training begins.
    • Data minimisation - whilst data minimisation may seem to conflict with the creation and training of models, which by their nature, need to be trained on large volumes of data, the DPC still encourages the following measures to minimise personal data processing: pseudonymisation or anonymisation; de-duplication; shorter collection periods; exclusion of under 18s' data; avoiding sources which are likely to contain special category data and avoiding sources such as dormant accounts.
    • Automated decision making - A human triggering the AI system, or an appeal process with human review, does not necessarily avoid automated decision making or ensure compliance with the obligation not to subject a data subject to a prohibited decision under Article 22 of the GDPR. 

    Organisations launching or deploying AI features in the EU (particularly where Ireland is its the lead authority) should consult the DPC Report for insight into how the DPC approaches regulatory supervision. 

    2. EU AI Office announces first AI Act investigations  

    On 14 October 2026, Commission Executive Vice-President Henna Virkkunen confirmed at the International AI Summit in Dublin that the AI Office had started enforcing the AI Act's rules for general-purpose AI models.  

    Under the AI Act, enforcement powers were effective from August 2026 and there had been more than 20 loss of control incidents documented in 2026 including two specific incidents over the summer where AI agents had escaped operating and testing environments and models had reportedly gained unauthorised access to three external organisations. The AI Office has made requests for information on security, external evaluation and monitoring of model use. 

    The cases are expected to test the scope of the AI Act and whether its research and testing exemption means it does not apply to incidents before a model is placed on the market. The Commission's reported position is that it will not enforce during testing, but that the Act becomes enforceable where a model escapes its testing environment and affects the EU market.  

    Organisations procuring or building on general-purpose AI models should expect closer regulatory scrutiny of model security and testing practices, and should consider seeking contractual assurances from providers on AI Act compliance. 

    EU | In case you missed it  

    1. On 21 October 2026, the Council of the EU agreed and adopted its negotiating position on the Digital Omnibus – this is the Commission's November 2025 proposal to simplify the EU digital rulebook, including the GDPR, the ePrivacy Directive and cyber incident reporting rules. The Council’s position: maintains the current definition of personal data; supports a principle-based definition of scientific research; introduces a new provision on pseudonymisation; supports the Commission’s proposal to extend breach notification deadlines but replaces the proposed single-entry point at EU level with a national entry point in each member state; introduces safeguards for the processing of special category data for biometric verification and development of AI systems; and maintains cookies rules under ePrivacy but provides additional purposes for where consent is not required. Trilogues can begin once the European Parliament agrees its own position. See more details here.
    2. On 16 October 2026, the European Commission published the fourth (and most extensive to date) version of the ERA Living Guidelines on the responsible use of Generative AI in research. The update adds new guidance on agentic AI and expands guidance on AI in peer review and evaluation of proposals, while retaining the existing practical approach under a revised structure. Research integrity principles of accountability, transparency and responsibility remain central, and researchers remain responsible for outputs regardless of the tasks delegated to AI agents. The next revision is planned for the third quarter of 2027. See more details here.
    3. On 29 October 2026, the six-month deadline for compliance with the Italian Garante's guidelines on tracking pixels in email communications (adopted on 17 April 2026) ended. The guidelines set out the framework and required transparency and consent requirements for use of tracking pixels. The use of tracking pixels and monitoring of user behaviour will constitute the access to terminal equipment pursuant to the Italian Privacy Code for which consent is generally required. The guidelines allow for consent for tracking pixels to be collected as part of consent collection for general marketing activities. The Garante has included compliance with the guidelines in its inspection plan. See more details here.
    4. On 9 October 2026, the EDPB reported that the Italian DPA had fined IQVIA, a company in a multinational group that carried out health data analytics and clinical research EUR 7,000,000 for unlawful processing of health data. It related to a database of health information relating to approximately one million patients whose data was used for studies including those commissioned by pharmaceutical companies. The Garante found that the data was not anonymous as claimed, since a persistent patient identifier combined with detailed attributes (such as year of birth, diagnoses and location data) allowed individuals to be singled out and potentially re-identified using reasonably available means. This fine is a reminder that truly anonymised data has a high threshold of anonymisation; any organisations considering using anonymised data, should ensure compliance with the EDPB guidelines on anonymisation. See more details here.
    5. On 16 October 2026, the EDPB reported that the Norwegian DPA had fined Xandr Inc. (now part of Microsoft) EUR 8,600,000. The fine was imposed for the processing of: Grindr users’ personal data (including IP addresses and advertising IDs) during real-time bidding without a legal basis; and sexual orientation data without consent. Xandr was found to be a joint controller and could not rely on Grindr's (invalid) consent. This decision showcases that in joint controller relationships, accountability cannot be outsourced and each joint controller must verify that the consent relied on is valid and covers its own processing activity. See more details here.
    6. On 13 October 2026, the EDPB adopted a Statement on smart glasses which addresses key data protection issues such as data protection designation, transparency for non-users/bystanders, ePrivacy and AI training which are still relevant to similar wearable devices. Manufacturers and providers of connected wearables should review recording indicators, default settings (including data minimisation and storage), AI training practices, controllership allocations and review how transparency information is provided. See more details here. 

    Authors: Rhiannon Webster, Partner; Alexandre Brazeau, Partner; Alexander Duisberg, Partner; Arnav Joshi, Partner; Nicolas Quoy, Partner; Shehana Cameron-Perera, Senior Associate; Tom Brookes, Senior Associate; Stacy Young, Associate.

    The information provided is not intended to be a comprehensive review of all developments in the law and practice, or to cover all aspects of those referred to.
    Readers should take legal advice before applying it to specific issues or transactions.