Newsletter #47 - 17 November 2026
NicFab Newsletter
Issue 47 | 17 November 2026
Privacy, Data Protection, AI, and Cybersecurity
Welcome to issue 47 of the weekly newsletter dedicated to privacy, data protection, artificial intelligence, cybersecurity and ethics. Every Tuesday you will find a curated selection of the most relevant news from the previous week, with a focus on European regulatory developments, case law, enforcement and technological innovation.
In this issue
- EDPS - EUROPEAN DATA PROTECTION SUPERVISOR
- COUNCIL OF THE EUROPEAN UNION
- EUROPEAN COMMISSION
- EUROPEAN PARLIAMENT
- GARANTE PRIVACY ITALY
- CNIL: FRENCH DATA PROTECTION AUTHORITY
- CASE LAW
- DIGITAL MARKETS & PLATFORM REGULATION
- EUROPEAN STANDARDIZATION
- INTERNATIONAL DEVELOPMENTS
- ARTIFICIAL INTELLIGENCE
- CYBERSECURITY
- TECH & INNOVATION
- SCIENTIFIC RESEARCH
- AI ACT IN A NUTSHELL: Part 47
- LEGAL PROMPTING
- PODCAST
- THE BOOK “Agentic AI”
- FROM THE NICFAB BLOG
- Featured events and meetings
- Conclusion
EDPS - EUROPEAN DATA PROTECTION SUPERVISOR
EU KIDS Act: EDPS Opinion 36/2026
On 12 November, the EDPS published Opinion 36/2026 on the proposed EU KIDS Act Regulation, presented by the Commission on 17 September, and explained it in a press release the following day. The Supervisor shares the proposal’s objectives, namely protecting children online and avoiding fragmentation of the single market, and welcomes its safety-by-design approach. It also endorses two technical choices: the ban on using age assurance and parental verification tools to identify, locate, track or profile users or to target them with advertising, and the reference to privacy-enhancing technologies such as zero-knowledge proofs, which allow users to prove they meet an age threshold without revealing their age or identity.
The recommendations are specific. Legislators should identify which features genuinely require age assurance and where safety by design for all users is the better option, complemented by reporting tools for minors, tools for guardians, and digital literacy initiatives. The EDPS proposes amending the GDPR to set a single age of consent for all information society services, avoiding duplicate checks and diverging national rules. Age assurance and parental verification duties should apply only once privacy-preserving solutions are effectively available across the Union; age estimation should be allowed only with adequate safeguards, with algorithmic age estimation and other intrusive practices expressly excluded. On supervision, the Opinion calls for a legal basis for cooperation between competent authorities, including the exchange of confidential information gathered in audits and investigations: the same request the EDPB put to the Commission this very week, as reported below.
EDPS — Opinion 36/2026 of 12 November
|
EDPS — press release of 13 November
TechSonar 2026-2027: six technology trends under watch
On 12 November, the EDPS published the 2026-2027 edition of TechSonar, the foresight exercise through which it identifies emerging technologies and assesses their possible effects on privacy and other fundamental rights. Six trends were selected: Internet of Bodies, AI integration protocols, AI-based employee monitoring, digital content provenance and authenticity, online age assurance technologies and post-quantum cryptography. Wojciech Wiewiórowski’s foreword focuses on trust in the age of AI and safeguarding authenticity. Two of the six trend reports intersect directly with this week’s files: content provenance, at the heart of the Commission’s guidelines on Article 50 of the AI Act, and age assurance, at the heart of the Opinion on the EU KIDS Act.
EDPS — TechSonar 2026-2027 of 12 November
COUNCIL OF THE EUROPEAN UNION
Dublin, 13 November: EU leaders’ first discussion on AI
At their informal meeting in Dublin, the Heads of State or Government held their first strategic discussion on artificial intelligence, along two lines: how to harness its potential for European competitiveness, and what effects it has on society, security and trust. On the first, joined by Arthur Mensch (Mistral AI), leaders addressed the mobilisation of public and private investment at scale and the related infrastructure, the strengthening of European capabilities by securing value chains and reducing critical dependencies, access to clean and affordable energy, and the uptake of AI by companies and public administrations. President Costa pointed to the savings and investments union and the so-called 28th regime, warning that catching up with global competitors will require significant additional funding and that the next multiannual budget, in particular the new European Competitiveness Fund, will play a decisive role.
On the second, with Nobel laureate Geoffrey Hinton, the discussion covered the protection of fundamental rights, including children’s rights, the integrity of democratic processes, the impact on the European social model and labour markets, market concentration and security challenges, cybersecurity included, as well as international cooperation on the safety of advanced AI. Costa presented the AI Act as the world’s first AI governance framework and announced that AI and jobs will be the central theme of the 2027 social summit. For data protection practitioners, the key point is that fundamental rights and child protection now sit on leaders’ agendas alongside competitiveness, the same week the EDPS issued its Opinion on the EU KIDS Act.
European Council — main results of 13 November
|
European Council — remarks by President Costa of 13 November
Cybersecurity Review package and Digital Networks Act: the Committee of the Regions’ opinion
On 12 November the Council circulated the opinion of the European Committee of the Regions (rapporteur Zdeněk Hřib, CZ/Greens) on the package comprising the proposed Regulation on ENISA, the European cybersecurity certification framework and ICT supply chain security (repealing Regulation (EU) 2019/881), the proposed Directive amending NIS2 to align it with the new Cybersecurity Act, and the proposed Regulation on digital networks, which amends, among others, Directive 2002/58/EC and repeals Regulation (EU) 2018/1971, Directive (EU) 2018/1972 and Decision No 243/2012/EU. The Committee asks for the proportionality test on supply chain measures to be anchored in the text of the Regulation, so as to avoid rip-and-replace obligations for whole technology stacks that local authorities could not bear without disrupting essential public services. It also calls for a dedicated envelope, within the digital window of the European Competitiveness Fund, for capacity building in local and regional authorities, and for territorial cohesion and subsidiarity to be respected in the future digital networks framework. On the data protection side, the amendment to the ePrivacy Directive carried by the Digital Networks Act deserves attention.
Council — doc. 15800/26 of 12 November
EDPS Opinion 33/2026 on the recognition of qualifications of third-country nationals
With a cover note of 11 November, the Council circulated to delegations EDPS Opinion 33/2026, received by the General Secretariat on 6 November, on the proposed Directive on the recognition of qualifications of third-country nationals (interinstitutional file 2026/0275 (COD)). The file sits at the intersection of the internal market, education, visas, asylum and migration, as the document’s classification codes indicate.
Council — doc. 15760/26 of 11 November
EDPS opinion on the European Social Security Pass (ESSPASS)
Also on 11 November, the Council circulated the EDPS opinion on the proposed Regulation on digitalising the coordination of social security systems, laying down a framework for the European Social Security Pass (ESSPASS) and amending Regulations (EC) No 883/2004, (EC) No 987/2009 and (EU) 2018/1724, the latter establishing the Single Digital Gateway (file 2026/0271 (COD)).
Council — doc. 15777/26 of 11 November
Working Party on Data Protection: the revision of Regulation (EU) 2018/1725
On 13 November, the Working Party on Data Protection met in Brussels, in 1+2+1 format, on the proposed Regulation amending Regulation (EU) 2018/1725 on the processing of personal data by Union institutions, bodies, offices and agencies. The core item was the Presidency compromise text (doc. 15174/26) with discussion note WK 16894/2026; the agenda also included a Commission update on data protection developments and the presentation of the incoming Lithuanian Presidency’s priorities and calendar, which signals where the file will land in the first half of 2027.
Council — doc. CM 4827/1/26 REV 1 of 12 November
EUROPEAN COMMISSION
Consumer Package: Digital Fairness Act and direct enforcement powers for the Commission
On 11 November, Executive Vice-President Virkkunen and Commissioner McGrath presented the consumer package, which includes the proposed Digital Fairness Act, a revised Consumer Protection Cooperation (CPC) Regulation, and a simplification directive with targeted amendments to the Consumer Rights Directive. The Digital Fairness Act bans dark patterns, such as false scarcity claims that push consumers into purchases; requires transparent pricing for in-app virtual currencies and dynamic pricing for cultural and sports events; introduces the option to switch off design features that encourage compulsive use; and bans gambling-like mechanics, such as pay-to-win, for minors. The proposal also restricts personalised advertising and pricing that target minors or exploit sensitive data, makes digital contracts and subscriptions easier to cancel, establishes a right to speak to a human in customer service and sets clear rights for consumers buying through an AI agent.
The CPC revision gives the Commission the power to investigate large-scale cases with a significant impact across the single market directly, with fines of up to 6% of worldwide annual turnover, speeds up coordination between national authorities and requires non-EU companies to appoint a legal representative in the Union. The proposals now move to the European Parliament and the Council. For data protection practitioners, the point to watch is the ban on personalisation based on sensitive data, which overlaps with Article 9 GDPR and will require explicit coordination between consumer authorities and supervisory authorities.
Commission — press release IP/26/2407 of 11 November
|
Commission — questions and answers QANDA/26/2408
|
Commission — remarks SPEECH/26/2412
AI Act: the implementing regulation on regulatory sandboxes
On 10 November, the Commission adopted Implementing Regulation (EU) 2026/2487, published in the Official Journal on 11 November and applicable from 1 December, laying down rules for the establishment, development, implementation, operation and supervision of the AI regulatory sandboxes provided for by the AI Act. The text governs the general conditions for access and participation, applications and selection, the sandbox plan, the written proof and exit report, follow-up activities, and the extension and suspension of projects. The plan agreed between the provider and the competent authority must indicate, among other things, whether the system involves the processing of personal data and the need to involve the competent data protection authority; the recitals specify that such authorities should be associated with the operation of the sandbox and supervise where personal data are processed for the development, training or testing of the system. Sandboxes set up by the AI Office give SMEs priority access. Exiting a sandbox does not confer a presumption of conformity, but market surveillance authorities and notified bodies should take the exit report into account to speed up conformity assessment, to a reasonable extent.
EUR-Lex — Implementing Regulation (EU) 2026/2487, OJ L of 11 November
EUROPEAN PARLIAMENT
AI and democracy: the resolution on the institutional aspects of artificial intelligence
On 12 November, the European Parliament adopted, by 405 votes to 133 with 60 abstentions, its resolution on the institutional aspects of artificial intelligence in the context of European integration (rapporteur Manolis Kefalogiannis, EPP, Greece). The text recognises AI’s potential to support EU lawmaking, from evidence-based policymaking to the simplification and evaluation of Union law, but insists that its use within the institutions remain human-centred and subject to human oversight, never replacing political responsibility or producing purely automated political decisions. MEPs call for European AI tools to be prioritised, starting with the institutions, to build strategic autonomy across the whole value chain, and for an international agreement on the most serious risks, modelled on nuclear non-proliferation regimes.
On the democratic side, the resolution calls for a systematic parliamentary assessment of the risks and opportunities of AI in electoral processes, reform of European electoral law, and the inclusion of new rights in the Charter of Fundamental Rights, including the protection of personal identity and likeness and algorithmic transparency. Finally, it proposes a budget-neutral parliamentary AI observatory to support committees in the ex post evaluation of legislation.
European Parliament — press release of 12 November
GARANTE PRIVACY ITALY
Meta fined €805 million: Android gallery accessed without specific consent
The Garante fined Meta Platforms Ireland €805 million for accessing, through the Facebook app for Android, photos, videos and metadata in Italian users’ galleries without specific consent or adequate information. The case concerns the feature “Get suggestions from your gallery while browsing Facebook”, available since 2023 and enabled by default: gallery access is requested through a single permission covering several features, preceded by a generic notice (“enable other features”) that does not let users understand they are also authorising the suggestions. Once granted, the permission allows the app to read the entire photo archive, as a rule, permanently, without users or the Authority being able to verify which data are actually used.
The provision infringed is Article 122 of the Italian Privacy Code, which transposes the ePrivacy Directive and protects the smartphone as part of the private sphere: any access to the information it holds requires free and informed consent, regardless of whether data are transmitted or harm occurs. The objection of lack of competence in favour of the Irish authority was rejected, since the ePrivacy regime has no one-stop shop. The investigation, opened in July 2025 following a report by the Centro Tutela Consumatori Utenti of Bolzano, did not cover any use of the content to train AI, which Meta denies; the Garante reserves separate inquiries into the new 2026 feature that processes gallery content in the cloud. Within 60 days, the feature must be disabled by default and consent collected through a separate, specific request.
Garante — press release of 14 November
RespiVirNet: the Garante did not suspend the data flows
The Garante clarifies that it did not order any suspension of RespiVirNet data flows: the decision was taken independently by the Regions and Autonomous Provinces. At the meeting of 9 November, convened by the Authority, the Ministry of Health acknowledged that the draft decree governing the surveillance system had significant shortcomings and decided to withdraw it and redraft it in line with the GDPR. Since 2012, the law has required every surveillance system to be governed by a ministerial decree issued after consulting the Garante; no rule prohibits epidemiological surveillance, provided it respects the safeguards for health data. Adoption and timing of the decree lie with the Ministry.
Garante — press release of 13 November
Newsletter No 556: overview of decisions
The newsletter of 12 November covers, besides the three decisions described below, a guide to safer online shopping ahead of Black Friday, a €25,500 fine on Techpartner for delays in answering a former employee’s access request and for the handling of his email accounts after the employment ended, an €8,000 fine for aggressive telemarketing, approval of new rules on digital delegation for the Electronic Health Record and the European data protection authorities’ educational portal for children and young people.
Garante — newsletter No 556 of 12 November
Matera radiology centre: patients’ reports exposed online
With decisions No 764 and No 754 of 7 October, announced in the newsletter of 12 November, the Garante fined the Centro Radiologico Madonna della Bruna in Matera €16,000 and MedStore Srl of Saronno €12,000; MedStore acted as processor for hosting, managing and maintaining the online reporting platform. An IDOR (Insecure Direct Object Reference) vulnerability allowed anyone, without authentication, to view other patients’ reports, including name, tax code, address and medical history, simply by changing the final numeric part of the URL received by email. The IT company had not applied the available security updates. Reports were also retained beyond the 45 days allowed because the automatic deletion feature malfunctioned. The case is a reminder that timely patching and continuous oversight of infrastructure are legal obligations, not optional good practice.
Garante — newsletter No 556 of 12 November
|
Garante — decision No 764 of 7 October
|
Garante — decision No 754 of 7 October
Ministry of Justice: unlawful processing of minors’ data
With decision No 811 of 20 October, announced in the newsletter of 12 November, the Garante declared unlawful the processing carried out by the Ministry of Justice in breach of Article 52 of the Italian Privacy Code. The Ministry’s online database published court decisions concerning minors without redacting data that allowed them to be identified, even indirectly: in about 22% of the decisions concerning minors, the home addresses of the parents or of the minor remained visible, and a search for “minori” and “residenti” across 3,702 judgments returned 806 results from 2023 onwards. For the Authority, the infringement cannot be considered minor, since judicial data may not, by law, be disseminated. The Ministry subsequently redacted the data.
Garante — newsletter No 556 of 12 November
|
Garante — decision No 811 of 20 October
Artificial intelligence in schools: the Garante’s favourable opinion on the Ministry’s draft decree
With decision No 787 of 20 October, announced in the newsletter of 12 November, the Garante issued a favourable opinion on the Ministry of Education and Merit’s draft decree providing initial guidance on the use of AI solutions for the benefit of students and teachers. The draft includes a technical annex setting out organisational and security measures for pseudonymising user accounts and the informed use of AI systems in schools. The Authority’s remarks are addressed to the Ministry, given its steering and coordinating role over schools: provide adequate support to schools, including through future regulatory action, in view of the rapid evolution of the tools; guide them in choosing among the solutions on the market, including by defining standards, requirements and models that providers should meet in order to operate in the school sector; and consider identifying standard AI services to be made available to schools and setting up an assistance service to support them in their dealings with providers. The opinion is usefully read alongside the AgID guidelines on AI in public administration, adopted the same week and reported below.
Garante — newsletter No 556 of 12 November
|
Garante — decision No 787 of 20 October
CNIL: FRENCH DATA PROTECTION AUTHORITY
CNIL: ten simplified sanctions against local authorities for cybersecurity failures
In November, the CNIL issued ten simplified sanctions against local authorities and one public body for breaches of Article 32 GDPR, closing the thematic inspection programme launched in 2025. Recurring shortcomings: no multi-factor authentication for administrator accounts and remote access; permanent access for external vendors, often through shared accounts and without case-by-case authorisation; incomplete logging or logs not analysed in good time; insufficient network segmentation; obsolete server operating systems without security updates; weak passwords sent in clear text or stored without confidentiality safeguards. These bodies process civil status records, social benefits, financial data and minors’ data: a footprint that explains the choice of the simplified procedure as a tool for serial enforcement on standardised technical gaps.
The CNIL at the Salon des maires 2026
From 24 to 26 November, the CNIL will attend the Salon des maires et des collectivités locales (Porte de Versailles) and will speak on 25 November at the AI Forum of the 108th Congress of Mayors on municipalities’ use of AI tools. The 2026 edition of the practical GDPR guide for local authorities is available, with a four-step action plan, practical sheets and guidance on handling breaches. A webinar on the GDPR for local authorities is scheduled for 8 December, with registration opening on 23 November.
EDPB: the final version of the legitimate interest guidelines
On 10 November, the EDPB published version 2.0, the final version, of Guidelines 1/2024 on processing based on Article 6(1)(f) GDPR, together with the report on the public consultation, which had gathered 80 contributions between October and November 2024, and a flowchart for the Legitimate Interest Assessment (LIA). The Board mainly took on board requests for clarification, while many critical observations, which it viewed as resting on misreadings of the GDPR or the guidelines, did not lead to substantive changes. Among the new elements: clarification of the criteria for an interest to be relevant, in particular that it must be present and effective, while maintaining that all three must be met; confirmation that both the necessity test and the balancing test must be passed; a paragraph on the “weight” of the interest in the balancing test, which may increase where the interest coincides with that of a wider community; more attention to the effects, including positive ones, of the processing on data subjects and to the link with the DPIA under Article 35; more detailed guidance on mitigating measures, with the controller having to show how they change the outcome of the balancing test; and, as good practice, the possibility for data subjects to obtain information on the balancing test upon request. The Board also clarifies that Article 6(1)(f) is not, in itself, Union law authorising automated decision-making under Article 22(2)(b), and refers to Opinion 28/2024 for AI models and updates the section on fraud prevention.
EDPB — Guidelines 1/2024, version 2.0 of 10 November
EDPB: South Korea adequacy review, final BCR-P, sixth coordinated action
At its plenary on 10 November, the EDPB adopted three acts. Statement 2/2026 on the first review of the adequacy decision for the Republic of Korea assesses the amendments to PIPA: a general right to withdraw consent, a right to request the suspension of processing, and a right to refuse an automated decision, including one taken by AI systems, or to obtain an explanation where it significantly affects the data subject’s rights or obligations. On onward transfers, the EDPB welcomes the new regime under Article 28-8 PIPA, which requires specific transfer tools as alternatives to consent, and agrees with the Commission’s proposal to review the decision every four years in consultation with the committee established under Article 93(1) GDPR.
Following public consultation, the final recommendations on Binding Corporate Rules for processors (BCR-P) were adopted; they now apply to all BCR-P holders, who will have to align either during the approval process or with their 2028 annual update. Finally, the sixth Coordinated Enforcement Framework (CEF) action, to be launched in 2027, will address automated decision-making, including AI-driven decision-making, in recruitment, focusing on the right not to be subject to decisions based solely on automated processing and the related right of access.
|
EDPB: onward transfers also under Council of Europe and Ibero-American model clauses
With Statement 3/2026, adopted on 10 November and announced on 13 November, the EDPB welcomes the update of the Commission’s FAQs on Standard Contractual Clauses (SCCs): the model contractual clauses developed by the Council of Europe and by the Ibero-American Data Protection Network may also be used for onward transfers of data received under the EU SCCs, subject to the conditions set out in the FAQs. The outcome results from joint work between the Board and the Commission, and the Statement offers practical guidance to organisations wishing to rely on these instruments. For those operating in settings where the Council of Europe or Ibero-American clauses are already in use, this opens an additional contractual route for onward transfers.
|
EDPB — Statement 3/2026 of 10 November
EDPB to the Commission: a legal basis is needed for cooperation between regulators
In a letter of 12 November to von der Leyen, Virkkunen and McGrath, EDPB Chair Anu Talus follows up on the high-level meeting held in Dublin in July and asks the Commission to consider a legislative proposal grounding in Union law cooperation and the exchange of information, including confidential information, between the authorities enforcing digital rules: data protection, consumer protection, competition, digital services, market surveillance, media. The Board notes that the GDPR and the Procedural Regulation (EU) 2025/2518 govern exchanges between data protection authorities, but not with other kinds of regulators, and cites cases in which an authority could receive information from a national regulator without being able to share it with its counterparts in other Member States. The example of pay-or-consent models, examined in parallel under the GDPR, the DSA, the DMA and consumer law, shows the risk of diverging assessments, duplicated investigations and tensions with the ne bis in idem principle.
The proposal should identify the authorities concerned, the purposes and conditions of the exchange, with safeguards on professional secrecy, confidentiality and procedural rights, and allow sharing between authorities in different Member States as well. In the Board’s view, there would be no need to reopen the individual acts of the digital framework: a horizontal basis would suffice, as happened with the Representative Actions Directive.
CASE LAW
Court of Justice of the EU, Case C-579/24, Austro-Mechana and AKM
12 November 2026 — Second Chamber — ECLI:EU:C:2026:971
Online content-sharing services and Article 17 of Directive (EU) 2019/790. Storing works uploaded by users on the provider’s servers, where technically required to communicate them to the public, is an act of reproduction within the meaning of Article 2 of Directive 2001/29/EC attributable to the provider, for which no authorisation separate from the one the provider must obtain for communication to the public is required. Uploading by users is also a reproduction, but it is covered by the authorisation obtained by the provider, without users having to obtain their own.
Court of Justice of the EU, Case C-826/24 P, Commission v Google and Alphabet (Google AdSense) — Opinion of Advocate General Kokott
12 November 2026 — ECLI:EU:C:2026:969
The Commission appeals against the General Court judgment of 18 September 2024 (T-334/19), which annulled the 2019 decision on abuse of a dominant position in online search advertising intermediation. The Advocate General considers the grounds concerning the exclusivity clause and the placement clause largely well founded: the General Court is said to have applied the Intel criteria, in particular the duration of contractual commitments, with an incorrect standard of review and of proof, unduly reversing the burden of proof to the Commission’s detriment. She proposes setting aside the judgment and referring the case back to the General Court. This is relevant to the characterisation of contractual practices of digital advertising platforms, an area where the DMA rules also apply.
We have no ECtHR judgments under Article 8 of the Convention concerning data protection or surveillance to report this week.
DIGITAL MARKETS & PLATFORM REGULATION
DSA: consultation on the update of the election guidelines
On 10 November, the Commission launched a consultation on updating its 2024 guidelines for very large online platforms and search engines (VLOPs and VLOSEs) on assessing and mitigating systemic risks to electoral processes under Article 35 DSA. The revision draws on two years of experience with the DSA and the Code of Conduct on Disinformation and responds to technological change, generative AI in particular, and to increasingly sophisticated manipulation techniques. The proposed measures aim to counter information manipulation, make recommender systems safer and strengthen the protection of candidates. Contributions can be submitted until 23 November.
Commission — news of 10 November
Microsoft outage revives the European digital sovereignty debate
A Microsoft outage lasting almost 24 hours and affecting users in 74 countries overlapped on Friday with the Dublin summit on artificial intelligence. Mistral CEO Arthur Mensch reminded leaders that only 20% of European digital services run on European infrastructure, proposing that spending now going to US and Chinese providers be reinvested in Europe; Macron wrote on X that European independence cannot be outsourced. Finnish MEP Aura Salla called for public procurement to favour European technology, German Social Democrat MP Johannes Schätzl pointed to the continuity of public services during incidents like this as the real measure of digital sovereignty, and Sebastian Toffaletti of the Digital SME Alliance read the episode as evidence of systemic concentration risk. It remains to be seen whether the debate will translate into concrete clauses in public tenders, such as exit strategies and multi-cloud requirements.
EUROPEAN STANDARDIZATION
CEN-CENELEC: draft CWA on telehealth adoption open for comment
On 9 November, CEN and CENELEC opened a 30-day public commenting phase, running until 9 December, on the draft CEN Workshop Agreement “Framework for Telehealth Adoption: Planning, Integration and Outcome Evaluation in Healthcare Systems”, agreed by participants in the workshop launched on 20 May 2026. The document sets out a framework of requirements for evaluating, planning and integrating telehealth services into existing care pathways, based on the Telehealth Evaluation Model developed in the Horizon project “Caring Nature”, with clinical, technological and regulatory eligibility criteria and organisational readiness. Cybersecurity risks are listed among those to be assessed, while the draft under consultation contains no explicit reference to the GDPR: a point on which the commenting phase gives DPOs and healthcare organisations an opportunity to weigh in.
CEN-CENELEC — news of 9 November
Cyber Resilience Act: delivery of part of the M/606 standards postponed
On 10 November, the Commission amended standardisation request M/606 in support of the Cyber Resilience Act, covering 41 horizontal and vertical standards, extending by three months the delivery deadline for all vertical standards and some horizontal ones. Horizontal standards provide a common framework and cross-cutting processes for CRA compliance, including vulnerability handling; vertical standards, specific to product categories, are intended to provide a presumption of conformity with the essential requirements. For manufacturers and compliance officers, the extension pushes back the point at which they can rely on harmonised standards to demonstrate compliance.
Commission — CRA standardisation page, amendment of 10 November
INTERNATIONAL DEVELOPMENTS
Australia: mandatory reporting and a possible “failure to prevent” offence after the OpenAI–Medicare case
On 10 November, Prime Minister Anthony Albanese released the recommendations of the rapid review, led by the Department of the Prime Minister and Cabinet, into the June breach of a Medicare statistics reporting portal run by Services Australia, in which an AI agent developed by OpenAI accessed data from that system and three other government systems. OpenAI will not be prosecuted, but the notification delay weighed on the political response: the company detected the incident in August and informed Canberra only in mid-September, according to Guardian Australia, with an email drafted with AI help.
The review recommends that the Attorney-General’s Department, the Australian Federal Police and Home Affairs consider penalties for egregious regulatory non-compliance and a new “failure to prevent” offence targeting those who develop models capable of causing harm without taking reasonable steps to prevent it. The government will set up an AI safety regulator and impose six obligations on providers: security controls across the development and deployment lifecycle, identification and management of serious risks, rapid mandatory incident notification, independent model evaluation, publication of safety frameworks, and continuous alignment with standards. Mandatory, time-bound reporting of agent behaviour diverging from developers’ intentions is also envisaged. It is a producer-liability logic built on notification, independent evaluation and an omission offence, unlike the ex ante, risk-tier approach of the AI Act.
Australian Government (PM&C) — review page, recommendations of 10 November
|
United Kingdom: ICO guidance on AI-assisted subject access requests
On 11 November, the ICO published brief guidance on handling subject access requests generated with AI tools. Using AI does not, in itself, make a request complex; it must be assessed like any other: length, unclear wording, or legal references are not enough to justify extending the time limit, which depends on genuine difficulties in locating, reviewing, or disclosing the information. The guidance also covers unclear requests, reasonable and proportionate searches, metadata, manifestly unfounded or excessive requests and the risk of fake documents: where a letter of authority is suspected of having been generated with AI, the ICO suggests contacting the person the request is about directly. The issue also concerns European controllers, who are facing requests increasingly drafted with AI assistants.
Ibero-American Network: standards on personal data and AI approved
On 12 November, the Ibero-American Data Protection Network (RIPD) unanimously approved the Ibero-American Standards on the use of personal data in artificial intelligence systems, presented as the first international instrument to address comprehensively, through an articulated body of provisions, the application of data protection to AI processing. The Standards create neither a parallel regime nor new general obligations, but specify how existing rules apply, following an approach based on risk, proportionality and technological neutrality. Their content includes: protection of data that systems infer about individuals; the principle that public availability of data does not, in itself, authorise its use to train AI systems; the right to know whether one’s data have been used in an AI system and, for automated decisions, to challenge their conclusions and obtain genuine human intervention; and specific guidance on generative and agentic AI, biometrics, emotion recognition and neurodata. In the same week, the EDPB recognised the use of the RIPD model clauses for onward transfers.
ARTIFICIAL INTELLIGENCE
AgID adopts guidelines on AI adoption in public administration
With Director-General Decision No 169 of 9 November, AgID, the Agency for Digital Italy, adopted its Guidelines on the adoption of artificial intelligence in public administration, provided for by the 2024-2026 Three-Year Plan for IT in Public Administration and issued under the procedure of Article 71 of the Digital Administration Code (CAD), following public consultation, the opinion of the Unified Conference and engagement with the Italian data protection authority. The Guidelines are addressed to the entities listed in Article 2(2) of the CAD and apply to all AI systems and models as defined in Article 3 of the AI Act and referred to in Law 132/2025, Article 20 of which designates AgID as one of the national authorities for artificial intelligence, responsible for promoting innovation and development. The document distinguishes mandatory requirements from recommendations using ISO/IEC Directives terminology and covers the whole journey: principles for adoption, roles along the AI value chain and risk-based classification of systems; an organisational model running from strategy and context analysis to use cases, governance, risk management and impact assessment, through to monitoring and continuous improvement; compliance of solutions, with responsibilities along the value chain, oversight measures and record-keeping; ethical governance; transparency and information duties; and training and skills development. For DPOs and advisers to public bodies, it becomes the national operational reference against which impact assessments, privacy notices and procurement contracts should be aligned.
AgID — Decision No 169 of 9 November
|
EU Commission: guidelines on the transparency obligations of Article 50 AI Act
On 12 November, Communication C/2026/6005 was published in the Official Journal, setting out the Commission’s guidelines on implementing the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689. The document examines the article’s four obligations: informing people who interact directly with an AI system, with the exceptions for obvious interaction and for law-enforcement purposes authorised by law; marking and detectability of synthetic generated or manipulated content, with the effectiveness, interoperability, robustness and reliability requirements for technical solutions; informing people exposed to emotion recognition and biometric categorisation systems; and the obligation for deployers of generative AI systems to disclose clearly deep fakes and AI-generated or manipulated text published to inform the public on matters of public interest, unless the text has undergone human review or editorial control with editorial responsibility or law-enforcement purposes authorised by law apply, with a special regime for deep fakes forming part of artistic, creative, satirical or fictional works. The guidelines also recall that providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the marking obligation under paragraph 2. The document also clarifies the personal scope, the exclusions (purely personal non-professional use, research and development, components released under free and open-source licences) and the links with prohibited practices, high-risk requirements, AI literacy and the general-purpose AI regime. To be read alongside Articles 13 and 14 GDPR: the scope of Article 50 is autonomous, but it overlaps on several fronts with data protection information duties.
EUR-Lex — Communication C/2026/6005, OJ C of 12 November
Finland and Estonia: the EU needs an AI incident response mechanism
In a letter sent ahead of the Dublin summit, Finland and Estonia call for EU-level arrangements to respond to serious AI incidents and for independent European capacity to evaluate the safety and security of frontier models, together with investment in capital and compute. The proposal builds on a Hungarian paper from October that envisages a network of national contact points for cases in which providers lose control of their models, and a pool of pre-vetted, security-cleared independent evaluators. In the background: this summer’s incidents involving AI agents going rogue and recent coding vulnerabilities exploitable by automated systems.
Virkkunen: AI chips only with “very trusted partners”
At Web Summit in Lisbon, Executive Vice-President Henna Virkkunen said that on AI semiconductors, quantum, cybersecurity and satellites, the EU should work only with very trusted partners. These were her first public remarks since Nvidia opened an internal investigation into an alleged attempt by Chinese intelligence to target CEO Jensen Huang through a relationship with a Megaspeed executive. Virkkunen did not describe Nvidia as untrustworthy and named the United States as the traditional closest partner on security and technology, while stressing the need to prevent technological leakage and keep industrial secrets in Europe, with the Chips Act as the lever to reduce dependence in the design and production of advanced chips.
RAISE Act: New York opens the registration portal for frontier developers
On Tuesday, New York opened the registration portal for large frontier model developers with annual revenues above $500 million, who must register by 1 January 2027, when the RAISE Act takes effect. Obligations include publishing safety frameworks, reporting critical safety incidents within 72 hours and public model transparency reports; fines reach $1 million for a first violation and $3 million for subsequent ones. Enforcement goes to the new Office of Digital Innovation, Governance, Integrity, and Trust (DIGIT), supported by the New York Center for AI Responsibility and Research at SUNY Binghamton in evaluating the frameworks. The US landscape is widening: New York’s law on chatbots for minors takes effect on 1 June 2027, California’s SB 53 has been in effect since January, and Illinois’ SB 315 starts in early 2027 with independent audits.
The Pentagon renames CDAO “Super Intelligence Office”
The Department of Defense, now the Department of War, has renamed the Chief Digital and Artificial Intelligence Office the Super Intelligence Office, implementing Executive Order 14434 of September, which requires federal agencies to use “Super Intelligence” and “SI” in non-statutory communications. Cameron Stanley remains at the helm as Chief Super Intelligence Officer, with Andrew Mapes as Principal Deputy; public resources move to the SI.mil domain, and the portfolio still includes the Maven Smart System and the War Data Platform. The change follows October’s announcement of the “Super Intelligence Force” chaired by Director of National Intelligence Jay Clayton and the Super Intelligence Task Force set up by the Department of Justice on 5 November to coordinate civil and criminal enforcement.
FPF: a roundtable series on agentic AI, privacy and PETs
In a post on 12 November, the Future of Privacy Forum’s Centre for Artificial Intelligence reported on the first of four virtual roundtables on agentic AI, held on 19 October within the Privacy-Enhancing Technologies Research Coordination Network funded by NSF and DOE, in collaboration with Working Paper. Bennett Hillenbrand (Working Paper) described an agent as an LLM able to call tools, arguing that agents reshape existing privacy risks more than they create new ones and that advanced PET adoption was already marginal before their spread. Andrew Gruen (Working Paper, FPF) presented a preliminary taxonomy of 24 agentic risks, intended as a basis for pre-deployment benchmarks. Jaspreet Bhatia (Google) questioned static yes/no permissions for probabilistic agents, citing consent fatigue, the lack of protocols for negotiating privacy terms between agents, and deletion requests that never reach agents already holding the data. Antigoni Polychroniadou (J.P. Morgan AI Research) mapped enterprise risks from agent identity and human approval through to execution integrity and liability. The next session is set for 15 December.
CYBERSECURITY
ENISA: the 2026 report on the state of cybersecurity in the Union
On 10 November, ENISA presented the 2026 edition of its biennial report on the state of cybersecurity in the Union, required by Article 18 NIS2. The threat level remains substantial, and the EU’s overall posture has improved: the EU Cybersecurity Index averages 71.53 out of 100, with high scores on supply chain management of essential and important entities (89.91) and weaker ones on CSIRT maturity (54.76) and national certification ecosystems (55.56). Executive Director Juhan Lepassaar points to frontier AI models and the incidents affecting refrigeration systems as evidence of operational gaps and of the need for timely information sharing.
The report sets out a “compass” with five directions: effective implementation, with a coherent framework for entities operating across Member States and legislation; comprehensive operational capacity in crises; preparedness and resilience of critical infrastructure, where 70% of organisations cite regulatory compliance as the main driver of investment and AI risks widening the gap between sectors; management of risks and dependencies across supply chains, also in light of the industrialisation of AI-driven vulnerability discovery; and long-term societal resilience through awareness, skills and education.
ENISA — press release of 10 November
|
TSMC: employee data breach through a cloud provider
TSMC confirmed in an SEC filing that personal data of current and former employees had surfaced online after an actor using the alias “SmithBianl” auctioned 6.9 GB of data attributed to the company on a forum, with bidding starting at 3 BTC (about $65,000). The sample posted on 8 November contained 499 rows from a dataset said to cover 46,000 employees, including national ID numbers, addresses, contact details and banking information. The company’s checks traced the data to a cloud service provider it uses, with no evidence of compromise of its own systems; affected staff are being notified, and a third-party firm has been engaged. It is the classic external processor scenario: the controller remains responsible for notification and for demonstrating the measures imposed on the provider, wherever the intrusion took place.
BleepingComputer — 15 November
FBI: the ARTEX framework industrialises attacks with AI agents
The FBI issued a FLASH advisory on the offensive use of ARTEX, a penetration-testing framework written in Go that orchestrates external LLMs (Claude, GPT, DeepSeek, GLM, Grok, Kimi) through agents that handle reconnaissance, exploitation, privilege escalation, and exfiltration. Since August 2026, several actors have used it against the financial, government, technology, defence and healthcare sectors, with some attack chains reaching privileged access in under three hours; the FBI also links it to the recent ConnectWise supply chain breach affecting Microsoft 365 environments. The original repository was removed from GitHub on 8 October, but forks, archived copies and a commercial edition, ARTEX PRO, survive; in early November, more than 1,600 publicly exposed consoles had been identified. The advisory includes indicators of compromise and command-and-control infrastructure, and MalwareMustDie has released detection signatures. The key point is the lower skill threshold now needed for complex attacks.
BleepingComputer — 15 November
Cursor: from the Fireworks AI breach to the trojanised Claude Code extension
Cursor has reconstructed the incident that, at the end of October, led to trojanised versions of the official Claude Code extension being distributed on its marketplace. The attacker used an access key issued to Fireworks AI and exposed in that vendor’s breach to read a bucket shared with it, containing 2025 coding agent traces from sessions run outside Privacy Mode. These included unredacted credentials of fewer than 0.01% of customers, used to pivot into internal SaaS services, archived non-production source repositories, customer support data and usage metrics. On 29 October, the attacker uploaded malicious versions 2.1.312 and 2.1.313 to the marketplace cache, which was exposed for about 11 hours to up to 5,000 users. The case is instructive on three fronts: retention of traces containing plaintext credentials, overly broad keys granted to third parties, and protection that depends on a mode being switched on rather than a default setting, at odds with the data minimisation principle.
Uber notifies thousands of drivers after the Fireworks AI breach
Uber is informing thousands of drivers that personal data and driving licence information were stolen following the October breach at its vendor Fireworks AI, which reported it to the company on 9 October. Notifications to state attorneys general show 2,567 people affected in Massachusetts, 3,430 in South Carolina and 1,873 in Rhode Island so far, with figures still pending in Illinois and notifications in California. Fireworks attributes the attack to a financially motivated group interested in LLM access tokens; the same breach also hit Notion. It is Uber’s second incident this year, after the one at Uber Freight. AI infrastructure providers are emerging as a shared point of vulnerability across very different sectors.
|
N-able: new zero-days exploited hours after PoC release
On 12 November, researcher Sean Bales released PoCs for two zero-days in N-able N-central, dubbed 4EverAble2Shell and Agent2Shell, tested against build 2026.4.1.26 HF2, which the vendor released on 10 November as an emergency patch. The first chain is an Apache Axis flaw (CVE-2026-90436) with an N-central privilege escalation that grants unauthenticated root access when the Legacy API is enabled; the second allows execution as SYSTEM on managed Windows devices via the agent. Attacks followed within hours: at least 21 servers compromised for cryptomining, 19 of them already patched. Bales has broken off coordination with the vendor, criticising its patching timelines; N-able has taken hosted environments offline and is restricting features and integrations. For managed service providers, the risk is twofold, since every compromised console opens access to clients’ systems.
BleepingComputer — 12 November
Apache HTTP Server: critical RCE in HTTP/2 (CVE-2026-119618)
Apache has fixed CVE-2026-119618, a race condition in the assign_headers() function of mod_http2 that can corrupt the heap and allow code execution with the privileges of the web server process. Affected versions are HTTP Server 2.4.55 to 2.4.70 and standalone mod_h2 2.0.8 to 2.0.42. The flaw, reported by Anthropic’s OSS Scanner and Google Cloud’s Vulnerability Research team, has no official CVSS score yet but is rated critical. According to Wiz telemetry, 64% of cloud environments host vulnerable resources, and about half expose them to the internet, regardless of whether HTTP/2, disabled by default, is enabled. No public exploits are known; watchTowr advises patching immediately or disabling HTTP/2. The same release fixes CVE-2026-106023 (mod_md) and CVE-2026-109163 (mod_vhost_alias).
BleepingComputer — 12 November
FBI: North Korea stole over $400 million in crypto with deepfake video calls
The FBI attributes a social engineering campaign dubbed FaceStealer to North Korea-linked actors; over eighteen months, it took more than $400 million in cryptocurrency from targets in the Web3 ecosystem. The Department of Justice seized the cloud infrastructure used and over $1.2 billion attributable to the TraderTraitor cluster following the Binance hack. The scheme, overlapping with the GhostCall campaign described by Kaspersky in October 2025 and attributed to BlueNoroff, uses compromised accounts of earlier victims to invite targets to video meetings, plays recordings of known contacts or deepfakes of industry figures to build trust, simulates an audio problem and lures victims into installing malware through a ClickFix-style decoy. Out-of-band verification processes for financial operations and credential sharing can no longer rely on visually recognising the other party.
Is AI making us worse security professionals?
A reflection in SecurityWeek starts from a minor episode, a colleague asking for presentation content as if querying a chatbot, to argue that pervasive AI use is eroding capabilities built over decades: the creativity to spot anomalies, humour as a communication tool, the patience to build an argument. It is not an anti-technology stance but a warning about skill atrophy, worth weighing when designing SOC workflows and training.
TECH & INNOVATION
Apple fixes iMessage, Brotli decoder and Bluetooth pairing flaws
Apple has released updates fixing two flaws in iMessage group chats: CVE-2026-113961 allows a remote attacker to join a group chat and receive its messages, while CVE-2026-113960 lets a user added to a group read messages sent before they joined. The same cycle closes CVE-2026-111497, an out-of-bounds write in Google’s Brotli decoder library, which CISA added to its Known Exploited Vulnerabilities catalogue on 27 October. The iOS and iPadOS 27.2.1 advisory covers all three flaws, whereas the notes for iOS 18.7.12 and watchOS 26.6.1 do not mention them, and Apple has not clarified whether those devices are vulnerable. All updates also fix Bluetooth and AirPlay flaws, including two that allow pairing without user interaction by a nearby malicious device and two that allow the limit on pairing code attempts to be bypassed.
SCIENTIFIC RESEARCH
Selection of the most relevant papers of the week from arXiv on AI, Machine Learning and Privacy
Auditing and Provenance of Fine-Tuned Models
Reconstructing LLM Lineage Trees via Fine-Tuning Path Inference
Submitted on 6 November 2026; announced on 9 November 2026.
The paper tackles a concrete problem: the provenance of models derived from LLMs currently rests on unverifiable self-declarations. The authors propose a method to infer not only the relationship between models but also the direction of derivation from parent to child. This is useful for checking licence compliance of derivative models, tracing inherited vulnerabilities, and auditing outsourced fine-tuning without relying on the provider’s statements.
ReplayAudit: Auditing Outsourced LLM Fine-Tuning via Sealed Records and Selective Replay
Submitted on 6 November 2026; announced on 9 November 2026.
A system for verifying outsourced fine-tuning based on sealed records and selective replay of intervals from sealed checkpoints, with telemetry-guided rather than uniform sampling. The goal is to reduce verification costs while retaining guarantees that the provider actually performed the agreed training. A useful technical reference when negotiating auditability clauses in AI outsourcing contracts.
Privacy Attacks on Federated Learning
Breaking the Privacy Promise: Gradient Inversion Attacks on Federated Low-Rank Adapted Vision Transformers
Submitted on 6 November 2026; announced on 9 November 2026.
Combining PEFT methods with federated learning, adopted to cut computational and communication costs, is often assumed to be safer for privacy as well. The authors show that gradient inversion attacks still work against LoRA-adapted Vision Transformers in federated settings, allowing training data to be reconstructed. Worth bearing in mind in DPIAs that justify federated architectures on the assumption that parameter efficiency also reduces privacy risk.
Consensus-Constrained Dual-Federated Optimization with Client-Side Dynamic Weighting for Distributed Android Device Threat Defense
Submitted on 6 November 2026; announced on 9 November 2026.
A federated learning proposal with consensus constraints and client-side dynamic weighting for threat detection on Android devices, addressing instability from partial participation and uneven local model quality. Interesting as a case in which federated learning is explicitly motivated by the refusal to collect device telemetry centrally.
Differential Privacy
On Differentially Private Mechanisms for Linear Regression
Submitted on 5 November 2026; announced on 9 November 2026.
A theoretical analysis of differentially private least-squares linear regression in both batch and continual observation settings, where the mechanism must release updated parameter estimates at each new data point in the stream. The continual observation model is the relevant one for dashboards and online statistics, where one-off analyses do not reflect the actual release pattern.
AI Agents and Authorisation Boundaries
Authorization Drift in LLM Agents: Evaluating Boundary Compliance in Multi-Step Tool Use
Submitted on 6 November 2026; announced on 9 November 2026.
The authors formalise “authorization drift”: LLM agents that, while pursuing legitimate goals and with no adversary involved, overstep authorisation boundaries on data, tools, scope and time. They release a benchmark of 648 samples across seven domains, each pairing a compliant path with an attractive unauthorised shortcut. Useful for anyone who has to define and document the operational limits of autonomous agents, including for the human oversight required by Article 14 of the AI Act.
Detection and Explainability
PHISHREV+: An ASP-Based Post-Hoc Defeasible Machine Learning Framework for Phishing Detection
Submitted on 6 November 2026; announced on 9 November 2026.
A framework that adds defeasible reasoning in Answer Set Programming to machine learning classifiers, so that new contextual evidence can be integrated after training without retraining the model. It addresses a known limitation: once the decision boundary is fixed, the classifier cannot absorb new signals. An interesting neuro-symbolic route to detection systems that stay up to date and are explainable.
MotionBench-XAI: Benchmarking What Shapes Shapley Explanations in Spatiotemporal Machine Learning
Submitted on 5 November 2026; announced on 9 November 2026.
A benchmark to understand what drives Shapley-value explanations on spatiotemporal data, with a use case on assessing Parkinson’s severity from motion capture data, separating spatial contributions (joints) from temporal ones (time windows). Relevant wherever explainability on health and biometric data is a requirement, not an option.
AI ACT IN A NUTSHELL: Part 47
Article 51 - Classification of general-purpose AI models as general-purpose AI models with systemic risk
Having examined Article 50 in Part 46 on the transparency obligations for providers and deployers of certain AI systems, we now turn to the core rules on general-purpose AI models in Article 51, which establishes when such models acquire the far more demanding regulatory status of models “with systemic risk”.
The provision sets out two alternative routes to this classification. The first (paragraph 1, point (a)) rests on a technical criterion: the model has “high impact capabilities”, evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks. The second (paragraph 1, point (b)) rests on a decision of the Commission, taken ex officio or following a qualified alert from the scientific panel, where the model has capabilities or an impact equivalent to those under point (a), having regard to the criteria set out in Annex XIII.
Paragraph 2 introduces a presumption of considerable practical significance: a model is presumed to have high impact capabilities when the cumulative amount of computation used for its training, measured in floating point operations (FLOPs), is greater than 10²⁵. The threshold acts as a proxy for the model’s power and potential systemic impact, but it does not exhaust the possibilities: a model below the threshold may still be classified as systemic through the decision-based route of point (b).
Paragraph 3 empowers the Commission to adopt delegated acts, in accordance with Article 97, to amend the thresholds in paragraphs 1 and 2 and to supplement benchmarks and indicators in light of technological developments, such as algorithmic improvements or increased hardware efficiency, so that the thresholds reflect the state of the art.
Practical implications
For providers of general-purpose AI models, systemic-risk status is a decisive turning point, because it triggers a regime of enhanced obligations. In compliance terms, this first requires precise technical monitoring of the computation used for training: the 10²⁵ FLOP threshold is not an abstract figure but a parameter to be measured, documented and tracked throughout the model’s development lifecycle.
Organisations developing models close to the threshold should put in place internal procedures for continuous self-assessment, with verifiable metrics and technical documentation that can demonstrate the computation used. It is also advisable to monitor regulatory developments: the Commission may amend thresholds and indicators through delegated acts, so the classification of a model that is below the threshold today should be reviewed periodically.
For DPOs, compliance officers and legal advisers, the second classification route is also relevant. Annex XIII lists qualitative and quantitative criteria: the number of parameters, the quality or size of the data set, the amount of computation used for training, the input and output modalities of the model, benchmarks and evaluations of its capabilities, its impact on the internal market due to its reach (presumed when it has been made available to at least 10,000 registered business users established in the Union), and the number of registered end-users. This calls for an analysis that is not merely arithmetical. Finally, the procedural dynamic of the scientific panel’s qualified alert, which can trigger Commission action, must be taken into account: providers should be ready to engage technically with the authority, with documentation, evidence and arguments supporting their position.
Legislative sources:
Next installment
In Part 48, we turn to Article 52 - Procedure.
LEGAL PROMPTING
A new series of the Legal Prompting column is in development, with in-depth installments on concrete use cases. The first ten-part cycle built, step by step, a method for the professional use of language models in legal practice: from prompt structure to the hierarchy of sources, from output verification to ethical and regulatory obligations. All installments remain available in the newsletter archive.
Legal Prompting: the new frontier of AI in the legal field
PODCAST
NicFab Podcast episodes, including those on Legal Prompting, remain available in the archive.
Listen →
THE BOOK “Agentic AI”
Agentic AI — Technical Autonomy, Human Responsibility, and Legal Governance
My book Agentic AI — Technical Autonomy, Human Responsibility, and Legal Governance (first edition, July 2026) is available, with chapter 16 by Valentina Grazia Sapuppo on technical standardisation for agentic systems, and a foreword by Antonino Caffo.
Artificial intelligence systems no longer merely generate content: they act, send, modify, and commit. The book starts there and offers a methodological answer to the three questions every organisation should ask when it delegates the capacity to act to a system: what limits to set before the action, what evidence to retain during it, and who must be able to answer for it afterwards. The thesis running through the whole work is unambiguous: technical autonomy does not create legal autonomy.
Working through the AI Act, the GDPR and the European digital regulatory ecosystem, the volume builds original categories — graduated agenticity, the chain of attribution of agentic action, evidential accountability — and hands them over to an operational instrument, the seven-column matrix, ultimately rendering them executable in code. Eighteen chapters in four parts (foundations, risks and legal characterisation, responsibilities and evidence, the operational governance model), plus the Conclusions and an apparatus designed for professional use: the reusable matrix template (Appendix A), two original Python tools reproduced in full and working (Appendix B), an annotated glossary, a table of sources and an analytical index.
It is a book for lawyers, DPOs, compliance officers and anyone who has to govern — rather than submit to — the delegation of action to machines.
The volume is available in paperback (310 pages) and in digital EPUB format, in Italian and English (Agentic AI — Autonomia tecnica, responsabilità umana e governance giuridica).
All the information →
FROM THE NICFAB BLOG
Marking, Provenance, Authenticity: Three Different Things
13 November 2026
What an AI label, a watermark and signed metadata mean, what Article 50 of the AI Act requires, what data they can reveal and why the absence of a mark proves nothing.
Lagrange 2: End-to-End Encrypted Sync for the Small Web
12 November 2026
Lagrange 2 and its end-to-end encrypted sync: the advantages of the Gemini protocol, how sync works, and a data protection reading of its design and limits.
NIS2 and management bodies: approving, overseeing, answering
11 November 2026
Management bodies and NIS2 in Italy: Article 23 of Legislative Decree 138/2024, documents to be approved, delegation, training, information on incidents and liability of natural persons.
NIS and GDPR in the face of an incident: the decisions to take before the crisis
10 November 2026
The incidents of October and November 2026 read through NIS and GDPR: where the two regimes meet, which technical details change the legal qualification, the role of artificial intelligence in attacks, prevention and compliance management, internal organisation and the decisions to take before the crisis.
Featured events and meetings
Grand Chamber judgment in Case C-698/23 P, EDPS v Parliament and Council (Europol Regulation) (scheduled: 17 November 2026)
Court of Justice of the European Union |
Apply AI Summit (scheduled: 17 November 2026)
European Commission |
Advocate General’s Opinion in Case C-643/25, Verbraucherzentrale Bundesverband (“free” digital service) (scheduled: 19 November 2026)
Court of Justice of the European Union |
EYCS Council: conclusions on skills and AI and on young people and the information environment (scheduled: 19 November 2026)
Council of the European Union |
Consultation on the update of the election guidelines under Article 35 DSA (deadline: 23 November 2026)
European Commission |
Salon des maires et des collectivités locales 2026, with the CNIL speaking at the AI Forum on 25 November (scheduled: 24-26 November 2026)
CNIL |
Application of Implementing Regulation (EU) 2026/2487 on AI regulatory sandboxes (scheduled: 1 December 2026)
European Commission |
Telecommunications Council (scheduled: 1 December 2026)
Council of the European Union |
Journée RGPD 2026 in Poitiers (scheduled: 1 December 2026)
CNIL |
Call for evidence on advertising repositories under Article 39 DSA (deadline: 2 December 2026)
European Commission |
Transitional deadline for marking synthetic content under Article 50(2) AI Act (systems placed on the market before 2 August 2026) (deadline: 2 December 2026)
European Commission |
Hearing in Case C-730/25, Vinted (scheduled: 2 December 2026)
Court of Justice of the European Union |
Feedback on the proposed EU KIDS Act Regulation (deadline: 3 December 2026)
European Commission |
Safer Internet Forum 2026 (scheduled: 3 December 2026)
European Commission |
Webinar: the GDPR applied to local authorities (scheduled: 8 December 2026)
CNIL |
Public comments on the draft CWA “Framework for Telehealth Adoption” (deadline: 9 December 2026)
CEN-CENELEC |
Product Cybersecurity: Making CRA a Reality (scheduled: 11 December 2026)
ENISA |
Plenary vote on the Child Sexual Abuse Directive (indicative date) (scheduled: 14 December 2026)
European Parliament |
Conclusion
Two documents from this week, read together, say the same thing from different positions. In its Opinion on the EU KIDS Act, the EDPS calls for a legal basis allowing competent authorities to cooperate and exchange confidential information gathered in audits and investigations. On the same day, the EDPB writes to the Commission asking for a legislative proposal that would enable information sharing between regulators with different mandates, across borders too, citing pay-or-consent models examined in parallel under the GDPR, the DSA, the DMA and consumer law. In my view, this is the most significant signal of the week: data protection authorities are openly stating that the current architecture of digital enforcement does not allow them to work on the same facts with other regulators.
Over the same days, the Commission proposes, through the Digital Fairness Act and the CPC revision, direct investigative powers and fines of up to 6% of worldwide turnover; Parliament calls for human oversight of AI within the institutions and for new rights in the Charter; and the leaders meeting in Dublin address AI at their level for the first time, placing children’s rights and democratic integrity alongside competitiveness. The implementing regulation on AI sandboxes provides for data protection authorities to be involved in projects processing personal data, and in Italy, AgID adopts its guidelines on AI in public administration after engaging with the Garante. ENISA, for its part, identifies timely information sharing as one of the most evident operational gaps. Instruments and fora are multiplying; what is missing, in my opinion, is the channel connecting them.
The week’s incidents show why that channel is needed. A Fireworks AI key opens a bucket of session traces containing plaintext credentials, and from there the attack reaches Cursor, while the same breach spreads to Uber and Notion. TSMC traces its employees’ data to a cloud provider. The same facts give rise to notification duties, security issues under Article 32 GDPR, contractual liability questions and, prospectively, product cybersecurity concerns: all within the remit of different authorities which, according to the EDPB’s letter, currently lack a clear basis for sharing what they know.
National responses are moving in different directions. The Garante fines Meta €805 million under Article 122 of the Italian Privacy Code, recalling that ePrivacy has no one-stop-shop; the CNIL closes an inspection cycle with ten simplified sanctions on standardised security gaps; Australia considers a “failure to prevent” offence for model developers; New York activates the RAISE Act portal with 72-hour reporting of critical incidents. My cautious prediction is that the next battleground will not be the quantity of rules but the ability of authorities to act in a coordinated way on the same facts: the EDPB’s letter is, in my view, the first explicit step in that direction.
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.