Newsletter #40 - 29 September 2026
NicFab Newsletter
Issue 40 | 29 September 2026
Privacy, Data Protection, AI and Cybersecurity
Welcome to issue 40 of the weekly newsletter on 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, focusing on European regulatory developments, case law, enforcement and technological innovation.
In this issue
- EDPB: EUROPEAN DATA PROTECTION BOARD
- COUNCIL OF THE EUROPEAN UNION
- EUROPEAN COMMISSION
- EUROPEAN PARLIAMENT
- GARANTE PRIVACY ITALY
- CNIL: FRENCH DATA PROTECTION AUTHORITY
- CASE LAW
- DIGITAL MARKETS & PLATFORM REGULATION
- INTERNATIONAL DEVELOPMENTS
- ARTIFICIAL INTELLIGENCE
- CYBERSECURITY
- TECH & INNOVATION
- SCIENTIFIC RESEARCH
- AI ACT IN A NUTSHELL: Part 40
- LEGAL PROMPTING
- PODCAST
- THE BOOK “Agentic AI”
- FROM THE NICFAB BLOG
- Featured events and meetings
- Conclusion
EDPB: EUROPEAN DATA PROTECTION BOARD
Google: EUR 403 million from the Irish DPC over location data
On 21 September 2026, the Irish Data Protection Commission announced its final decision in the own-volition inquiry into Google Ireland Limited that it opened in February 2020 as lead supervisory authority. The inquiry followed complaints from several European consumer organizations, including BEUC. The EDPB relayed the decision on 23 September.
The inquiry concerned the processing of location data in three features — “Web & App Activity”, “Location History” and “Location Accuracy” — between 25 May 2018 and 4 February 2020. The DPC found infringements on four fronts: the lawfulness and fairness of processing in Web & App Activity and Location History; accountability, because Google could not demonstrate compliance with the principles of lawfulness, fairness and transparency in Location Accuracy; transparency, for all three features; and the retention of location data in Web & App Activity and Location History. The EDPB summary refers to Articles 5, 6, 12 and 13 GDPR.
The administrative fines total EUR 403 million, together with an order to bring the processing into compliance within six months. The DPC will publish the full decision later; until then, the breakdown of the amount across the individual infringements is unknown.
DPC — press release of 21 September
|
EDPB — summary of 23 September
Fines and other corrective powers: Guidelines 04/2026 and the final DSA–GDPR guidelines
On 21 September, the EDPB announced the outcome of its plenary of 17 September 2026, whose agenda was reported in issue 39. The Board adopted Guidelines 04/2026 on the application of the power to impose administrative fines in relation to other corrective powers under the GDPR, and the final version of Guidelines 3/2025 on the interplay between the DSA and the GDPR.
Guidelines 04/2026 set out a five-step method for supervisory authorities. They must first check that the infringement can lead to a fine under the GDPR or national law; then identify who may be fined, controller or processor, depending on who is bound by the provision infringed; establish intent or negligence, a precondition for a fine; weigh aggravating and mitigating factors — a minor infringement will generally not lead to a fine and a reprimand may be issued instead, while for other infringements there is a strong presumption in favor of a fine; and, finally, check that the measure is effective, proportionate and dissuasive, with the possibility of departing from the standard approach on stated grounds.
The document also maps the full range of corrective powers (warnings, reprimands, orders, limitations including processing bans, withdrawal of certification) and provides 14 practical examples. It replaces the Article 29 Working Party guidelines on applying fines and complements the guidelines already adopted on calculating the amount. It is open for public consultation until 13 November 2026, so it is not yet final.
Guidelines 3/2025, by contrast, reach their final version after public consultation. They support the consistent application of both acts, in particular where DSA provisions concern the processing of personal data by intermediary service providers and refer to GDPR concepts.
|
|
|
AEPD: EUR 100,000 for Securitas Direct over a premium-rate number for data subject rights
On 22 September, the EDPB published a summary of a decision of the Agencia Española de Protección de Datos dated 1 February 2023. The news is therefore recent; the decision is not.
Following a complaint from a consumer association, the AEPD examined Securitas Direct’s video-surveillance signs, which referred data subjects to a chargeable 902 number to exercise their rights of access and objection. The authority found that the practice breached the requirement that rights be exercisable free of charge and could deter data subjects from exercising them. The availability of free channels on the website did not cure the infringement, because the sign referred specifically to that number.
The AEPD found an infringement of Article 12(2) GDPR, which requires controllers to facilitate the exercise of the rights under Articles 15 to 22, subject to the penalties under Article 83(5)(b). In addition to the EUR 100,000 fine, it ordered the signs to be replaced within twelve months.
EDPB — summary of 22 September
IMY: request for an EDPB opinion on Article 10 GDPR
On 23 September 2026, the Swedish authority IMY announced that it had asked the EDPB for an opinion clarifying which data fall within the protection of Article 10 GDPR on personal data relating to criminal convictions and offenses, and in particular whether the provision also covers data on offenses for which no legal proceedings have been initiated. The request forms part of the review of its own position that IMY launched in December 2025: its public consultation and exchanges with other data protection authorities revealed diverging interpretations across Member States.
The EDPB is to adopt its opinion within eight weeks, extendable by a further six. This is a request for clarification: the Board has not yet adopted its position.
COUNCIL OF THE EUROPEAN UNION
Cybersecurity incidents: the Court of Auditors’ special report before the Council
On 21 September 2026, the European Court of Auditors published Special Report No 19/2026 on detecting and responding to cybersecurity incidents. The title already sums up the assessment: the EU cooperation framework is progressing but is only partially effective because of implementation delays and limited information sharing.
In note 13547/26 of 24 September, the General Secretariat proposed that COREPER entrust the Horizontal Working Party on Cyber Issues with examining the report and invite the Court’s representatives to present its findings and answer delegations’ questions, in line with the procedure set out in the Council conclusions in doc. 7515/00 + COR 1. The note launches the Council’s examination; it does not anticipate its outcome.
Council of the EU — doc. 13547/26
Corrigendum to the German version of Regulation (EU) 2026/1744
In document 13206/26 of 22 September 2026, the Council Legal Service circulated a draft corrigendum concerning only the German version of Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), which amends Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230.
The correction concerns Article 1, point 8, which inserts new paragraph 1b into Article 6 of the AI Act on AI systems whose failure or malfunctioning would endanger health and safety, which are to be considered safety components. In the German text, the wording “Unbeschadet des Absatzes 1a” is replaced by “Ungeachtet des Absatzes 1a”, consistent with the meaning of the Italian version (“In deroga al paragrafo 1 bis”). This is the procedure for obvious errors in a single language version; Member States have three days to submit comments.
Council of the EU — doc. 13206/26
EUROPEAN COMMISSION
Data centers: the common Union rating scheme
On 21 September 2026, the Commission adopted Delegated Regulation C(2026) 3472, which supplements Directive (EU) 2023/1791 on energy efficiency, amends Delegated Regulation (EU) 2024/1364 and establishes a common Union rating scheme for data centers.
According to the Commission’s press release, the scheme will cover individual data centers with a capacity above 500 kW and make actual resource use — energy and water — transparent, along with their contribution to the grid: waste-heat reuse, new clean-energy generation capacity, and flexibility. The first sustainability labels are expected in 2027, and a first review is planned by the end of 2028.
The delegated act is now subject to a two-month scrutiny period by the European Parliament and the Council, which may object but cannot amend the text. For operators of cloud and AI computing infrastructure, the scheme adds to the reporting obligations already laid down in the Directive.
European Commission — IP/26/1667
The report on the energy efficiency of data centers
In its report COM(2026) 500 of 21 September 2026, the Commission analyzes the data submitted to the European database in the first reporting period. In 2024, data centers in the EU consumed 68 TWh of electricity; the 2030 estimate is 114 TWh, or 3.2% of EU electricity demand.
A total of 770 data centers reported, around 36% of those estimated to be subject to the obligation. The report examines four indicators — power usage effectiveness (PUE), water usage effectiveness (WUE), energy reuse factor (ERF) and renewable energy factor (REF) — and finds, among other things, an average PUE of 1.36 and an average WUE of 0.58, with wide differences between Member States. The identified lines of action are simplifying reporting, a common rating scheme, a proposal for minimum performance standards, and an assessment of the feasibility of a carbon-neutral data center sector.
Minimum performance standards for data centers: call for evidence
With document Ares(2026)8908107 of 21 September 2026, the Commission launched a call for evidence for an initiative on minimum performance standards for data centers operating in the Union, accompanied by an impact assessment and a public consultation. The consultation closes on 14 December 2026, and indicative planning places the proposal in the second quarter of 2027. The thresholds have therefore not yet been set.
European Commission — Have your say, initiative on minimum performance standards
|
European Commission — IP/26/1667
European Health Data Space: three implementing regulations
On 21 and 22 September 2026, the Official Journal published three implementing regulations under Regulation (EU) 2025/327 on the European Health Data Space (EHDS).
Regulation (EU) 2026/2083 of 18 September governs the MyHealth@EU infrastructure for the cross-border exchange of health data: cybersecurity, technical and semantic interoperability, operations and service management, and the allocation of roles, with the national contact points for digital health as joint controllers and the Commission as processor. It applies from 26 March 2027.
Regulation (EU) 2026/2098 of 18 September lays down the minimum metadata elements that health data holders must provide to describe datasets available for secondary use (Article 77(4) EHDS). It applies from 26 March 2029.
Regulation (EU) 2026/2099 of 21 September concerns the interoperable mechanism for identifying and authenticating natural persons, health professionals and healthcare providers. Each Member State will define a set of “health attributes”, which may be issued on request to the European Digital Identity Wallet; health professionals must use at least assurance level substantial, rising to high from 26 March 2032. It applies from 26 March 2027, with certain provisions from 2029.
|
|
Children and AI: von der Leyen at the United Nations
On 22 September 2026, in New York, the Commission President spoke at the high-level dialogue “Protecting Children’s Rights facing Artificial Intelligence”, at the launch of the Coalition for Children’s Rights and Protection in the Age of Artificial Intelligence.
In her speech, she recalled the content of the EU KIDS Act proposal presented the previous week: no social media under 13, no personal accounts under 15 and, between 13 and 15, reduced accounts set up and supervised by parents, with a limit of 60 minutes a day. She identified the reversal of the burden of proof as the central element, placing on platforms the task of demonstrating that their services are safe for children. This is a political statement on a proposal still before the co-legislators.
European Commission — SPEECH/26/1942
EUROPEAN PARLIAMENT
No plenary session of the European Parliament took place during the reference period (the last one was held from 14 to 17 September), and no acts or communications of the LIBE, IMCO and ITRE committees relevant to this review were identified.
GARANTE PRIVACY ITALY
The news item of 21 September on the National Diabetes Registry refers to opinion no. 582 of 6 August 2026, already covered in
.
Platform workers: the Garante President’s interview
On 24 September 2026, the Garante’s website republished the interview Pasquale Stanzione gave to Il Sole 24 Ore on protecting workers managed by digital platforms. The President calls on the Government and Parliament to introduce stronger safeguards for the privacy of so-called “digital invisibles” and to connect the rules on platform work with data protection law, recalling the historical link between privacy and the Italian Workers’ Statute.
The points raised include: the position expressed by the Garante on a decree concerning artificial intelligence, namely extending the ban on the use of AI to all decisions with a significant impact on workers, not only dismissal; the inviolability of the sphere of trade union rights; clearer rules on sharing information with public authorities; and the risk of overlapping penalties with the National Labour Inspectorate. This position was expressed in an interview, not a collegial decision.
Garante — interview of 24 September
CNIL: FRENCH DATA PROTECTION AUTHORITY
Solocal Marketing Services: injunction closed, penalty payment not enforced
On 24 September 2026, the CNIL announced that, by deliberation of 17 September, its restricted committee had closed the injunction issued on 15 May 2025 against Solocal Marketing Services.
In 2025, the company had been fined EUR 900,000 for breaches of the French Postal and Electronic Communications Code (CPCE) and the GDPR concerning the collection and proof of consent for commercial prospecting by SMS and email. The data came from data brokers and publishers of prize-draw and product-testing websites (“primo-collectants”), whose collection forms were found misleading and unable to support free and unambiguous consent. In addition to the fine, the company was ordered to stop prospecting without valid consent within nine months, subject to a penalty payment of EUR 10,000 per day of delay.
Within the deadline, the company showed that it had introduced an automated analysis of its partners’ collection forms, supplemented by human review. The restricted committee found this adequate, if applied rigorously, and decided not to enforce the penalty payment. It nonetheless recalled that the company remains responsible for the validity of consents collected on its behalf: the existence of a control mechanism does not exclude liability for non-compliant forms.
Agenda of the plenary of 24 September
The CNIL’s agenda for 24 September 2026 included two draft opinions on decrees relating to the legal protection of adults: the first on information transmitted to fund court-appointed representatives and family-benefit delegates, the second on the creation of the “SI-Mandoline” processing system. It also scheduled a communication on health data, the renewal of BYCYB’s accreditation as monitoring body for the CISPE code of conduct, an opinion on a draft Conseil d’État decree amending Article R.128-6 of the French Commercial Code, and deliberation HAB-2026-002 authorizing CNIL agents to carry out verification missions. The agenda records scheduled business, not the content of the opinions.
CASE LAW
A selection of this week’s judgments and Advocates General’s Opinions relevant to digital law, data protection, privacy, platform regulation and artificial intelligence. The citations distinguish the date of the ruling from subsequent commentary.
During the reference period, no judgments or Advocates General’s Opinions of the Court of Justice of the European Union relevant to the subjects of this review were identified.
ECtHR, Meshcheryakov v. Ukraine, application no. 27003/16
24 September 2026 (deliberated on 3 September 2026) — Committee, Fifth Section
Violation of Article 8 ECHR on account of the interception of the applicant’s communications: the decisions authorizing the interception had been destroyed, and the Court could not establish that they resulted from proper judicial scrutiny; violation also of Article 6(1) on account of the length of the criminal proceedings.
DIGITAL MARKETS & PLATFORM REGULATION
DMA: summary of the non-compliance decision on Google Search published
On 25 September 2026, the Official Journal (C series) published the summary of the Commission decision of 23 July 2026 in case DMA.100193, which found that Alphabet does not comply with Article 6(5) of Regulation (EU) 2022/1925 in relation to Google Search.
According to the Commission, the search results page ranks several Alphabet services, distinct from the search engine, more favorably than similar third-party services: Google Maps, the units for products, flights, trains, lodging and places, and the translation, sports and games units. The decision orders Alphabet to end the non-compliance within 60 days of notification, imposes a fine of EUR 460 million and provides for periodic penalty payments in the event of non-compliance. Alphabet must identify measures ranging from extending the same prominence to third parties to removing the prominence given to its own services.
Official Journal — C/2026/4985
DSA: good practices on notifying suspected criminal offenses (Article 18)
On 25 September 2026, the European Commission published the document on good practices and examples under Article 18 DSA, adopted by the European Board for Digital Services. Article 18 requires hosting service providers to inform law enforcement or judicial authorities promptly when they become aware of information that gives rise to a suspicion of a criminal offense involving a threat to the life or safety of persons.
The document clarifies when a provider can be considered aware of the information, gives a non-exhaustive overview of the relevant offenses, sets out examples of indicators of suspicion and good practices for identifying the Member State concerned and transmitting the information, and addresses supervision. The Board’s Working Group 7 prepared it, taking into account a targeted consultation held from 18 November 2025 to 9 January 2026, and it includes the list of national contact points. For hosting providers, a notification involves disclosing personal data to the authorities; internal procedures should therefore document its grounds and content.
European Commission — good practices under Article 18 DSA
DSA: the Board’s 20th meeting and the reports under the Code of Conduct on Disinformation
On 24 September 2026, the European Board for Digital Services held its 20th meeting in Brussels, covering, among other things, child online safety initiatives, the role of the DSA in election integrity, and trusted flaggers; in its statement, the Board welcomed the attention given to minors in the EU KIDS Act proposal.
On the same day, the Commission reported on the second round of reports submitted by the signatories of the Code of Conduct on Disinformation, recognized as a code of conduct under Article 45 DSA since 1 July 2025. The reports from the large platforms, including Google, Meta, Microsoft and TikTok, cover the period 1 January–30 June 2026. With a full year of reporting now available, the Commission will open dialogues with all signatories that are very large online platforms or search engines on their levels of commitment.
European Board for Digital Services — statement
|
European Commission — Code of Conduct
DSA: the fifth roundtable with civil society and researchers
On 25 September 2026, the Commission reported on the fifth online roundtable on DSA implementation with civil society organizations and researchers, held on 23 September with around 60 organizations. The discussion focused on systemic risks and mitigation measures, platforms’ risk assessment reports, independent research methodologies and data access, and the Commission’s enforcement priorities: protection of minors, online marketplaces, AI-related risks, and the EU KIDS Act proposal. The Commission presented the key steps of the enforcement process and the avenues through which civil society and researchers can contribute. It is a forum for exchange, not an enforcement act.
Digital Networks Act: the EESC opinion in the Official Journal
On 23 September 2026, the Official Journal (C series) published the European Economic and Social Committee’s opinion on the proposal for a regulation on digital networks (COM(2026) 16), which also amends Directive 2002/58/EC (ePrivacy). The opinion, with Maurizio Mensi as rapporteur, was adopted at the plenary session of 17 June 2026. The EESC considers a regulation the appropriate instrument and calls for avoiding overlapping sector-specific rules where horizontal EU legislation — consumer protection, data protection, cybersecurity — already provides effective protection, as well as for close alignment with NIS2, the Cybersecurity Act and the CER Directive.
Official Journal — C/2026/4657
INTERNATIONAL DEVELOPMENTS
Control of frontier models: the call promoted by Finland and Norway
On 21 September 2026, the Office of the President of the Republic of Finland published “A Call for Control of Frontier AI Models”, signed by Finnish President Alexander Stubb and Norwegian Prime Minister Jonas Gahr Støre together with other leaders, including the President of the European Commission. As of 24 September, the signatories represented 28 countries, in addition to the Commission; the United States and China are not among them. The text remains open for further endorsements.
The declaration notes that advanced AI systems have circumvented testing safeguards, exploited vulnerabilities, and gained unauthorized access to real-world systems, and states that AI must remain under human direction, oversight, and control. It calls on companies to adopt transparent safety protocols, with mandatory pre-deployment testing and independent evaluation; on governments to develop common standards and share reports of serious safety incidents; and on UN Member States to explore creating an international institution that can set standards, enable verification, and convene states when capability thresholds are crossed. It is a statement of political direction and has no binding effect.
Office of the President of the Republic of Finland
ARTIFICIAL INTELLIGENCE
Schools and generative AI: two Commission reports
On 22 September 2026, the Commission presented two reports. “Friend or foe? Evidence from the use of generative artificial intelligence in learning and teaching” describes generative AI’s shift from innovation to a mainstream tool in education: its effects on long-term skills remain uncertain and depend on how it is used, while teachers need substantial initial and continuous professional development. “Digital education at school in Europe 2026”, by the Eurydice network, finds that nearly two-thirds of education systems have adopted strategies, guidance, or other frameworks on AI, although gaps in access and technical support at school level remain. For schools, introducing these tools also involves choices about how to process students’ and teachers’ data.
CYBERSECURITY
ENISA Threat Landscape 2026: dependencies widen the attack surface
On 22 September 2026, ENISA published its Threat Landscape 2026, covering incidents and events observed from 1 January to 31 December 2025. Ransomware remains the incident type with the greatest short-term impact, and public administration is the most targeted sector (32% of cases); 73% of targeted organizations are essential or important entities under NIS2.
Among unauthorized-access incidents for which an intrusion vector could be identified (5%), 60% exploited a vulnerability. The report highlights supply-chain and third-party attacks, often with large-scale effects, and expects AI models to be increasingly used to support malicious operations. Low-impact DDoS attacks account for 51% of recorded cases.
CERT-EU: critical vulnerability in F5 BIG-IP APM, already exploited
On 22 September 2026, CERT-EU published security advisory 2026-013 on a critical vulnerability in F5 BIG-IP APM (CVE-2026-94127, CVSS score 9.8): a heap-based buffer overflow that allows an unauthenticated attacker to execute code remotely. Versions 17.1.0–17.1.3, 17.5.0–17.5.1 and 21.1.0 are affected when configured with an access policy and an OAuth profile on a virtual server. The vendor has confirmed active exploitation.
CERT-EU recommends preserving forensic evidence, applying the hotfix immediately, looking for indicators of compromise — including repeated OAuth authentication failures followed by suspicious commands — and starting the incident response process if any are found. Where the fix cannot be applied immediately, F5 provides an iRule-based mitigation on request.
Miljödata: IMY fine for security shortcomings
On 22 September 2026, the Swedish authority IMY fined IT systems provider Miljödata SEK 1.8 million for infringing Article 32(1) GDPR. In August 2025, an attack had affected, according to the company, the data of 2.2 million people, including personal identity numbers, contact details and sensitive data on sick leave, rehabilitation and school incidents, later published on the darknet.
IMY found insufficient controls for installing new software and no automated, real-time monitoring to detect intrusions and suspicious activity. The affected customers include most Swedish municipalities, several regions, and state authorities; inquiries into two municipalities and one region are ongoing.
Europol: the EDEN data protection conference
On 24 September 2026, Europol reported on the 17th conference of its Data Protection Experts Network (EDEN), held in Lisbon on 22 and 23 September, titled “Do Not Cross: Protecting Citizens, Preserving Rights”. The focus was on how law enforcement authorities can use data and technology while protecting fundamental rights and the rule of law, with a focus on data protection and cybersecurity.
TECH & INNOVATION
Commission and EIB: the institutional investors pact for tech scale-ups
On 22 September 2026, the European Commission and the EIB Group launched the European Institutional Investors Pact (EIIP) at the TechEU Equity Summit: a voluntary framework to channel more long-term institutional capital into Europe’s technology scale-up ecosystem. Thirteen institutional investors have expressed their intention to invest, including through the EUR 15 billion European Tech Champions Initiative 2.0 and the Scaleup Europe Fund, which targets EUR 5 billion in commitments.
The Pact comprises a Policy Dialogue Forum, led by the Commission, and an Investment Platform, led by the EIB Group, to facilitate access to venture capital and growth equity opportunities. The Scaleup Europe Fund invests in strategic technology areas, including artificial intelligence. The initiative focuses on company financing and imposes no regulatory obligations.
SCIENTIFIC RESEARCH
A selection of the week’s most relevant arXiv papers on AI, machine learning and privacy. Temporal criterion: announcement on arXiv during 21–27 September 2026; the submission date, where different, is stated for each paper.
LLMs, fine-tuning and personal data leakage
Decoupling Knowledge and Privacy: Post-Task Self-Distillation Replay for LLM Continual Learning
Submitted on 31 August 2026; announced on 25 September 2026.
The paper addresses continual learning in language models, aiming to reduce the reproduction of sensitive content without losing knowledge acquired across successive tasks. The authors distinguish formal privacy guarantees, which characterize randomized mechanisms, from operational control over outputs, and propose SPARK: the learned behavior is first consolidated, then a targeted correction is applied to reduce the likelihood of generating annotated personal data.
Not All 4-bit Quantizers Are Equal: Deployment-Time Mitigation of PII Leakage in Fine-Tuned Small Language Models
Submitted on 2 August 2026; announced on 23 September 2026.
The authors show that, when small models trained on private data are compressed to 4 bits, what affects personal data leakage is not the bit width but whether a calibration corpus is used to tune rounding: on the main model, the two calibration-based methods reproduce none of the records planted in the tests, while the calibration-free format reproduces 5.3% of them. In my view, the choice of quantization method thus becomes a factor to document in the impact assessment.
SSP-Bench: A Hybrid Data Generation Framework for Safety, Security, and Privacy Evaluation
Submitted on 21 September 2026; announced on 23 September 2026.
The framework generates evaluation instances on demand for the safety, security, and privacy of language models, to overcome the limits of static benchmarks: score saturation, data contamination, and insensitivity to equivalent rephrasings. Across 24 models, the results indicate that static benchmarks can misrepresent model behavior. In my view, the issue is relevant to the evaluations of general-purpose AI models with systemic risk required by the AI Act.
Inference and reconstruction attacks on distributed data
Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning
Submitted on 24 September 2026; announced on 25 September 2026.
The paper presents TRACE, a gradient-inversion attack against distributed reinforcement-learning agents that exploits the correlation between successive time steps to reconstruct sequences of private observations and actions beyond what single-frame attacks allow. The result questions the assumption that keeping raw data on the device is, by itself, sufficient protection; the authors point to the need for sequence-aware privacy mechanisms.
Learning Defensive Policies against Diverse Inference Attacks for Smart Meter Privacy
Submitted on 21 September 2026; announced on 23 September 2026.
Against inference attacks that infer individual appliance use from smart meters’ aggregate consumption (NILM), the authors propose battery-based load-shaping policies, learned through reinforcement learning, that inject realistic but misleading signatures. On two real-world datasets and six attackers unseen during training, the defense substantially reduces inference accuracy.
Hiding in Plain Sight: A Diffusion-based Mitigation of Geolocation Privacy Leakage in Vision-Language Models
Submitted on 18 September 2026; announced on 21 September 2026.
The study shows that multimodal reasoning models can infer users’ locations from shared photos by exploiting cues such as architecture, vegetation, and lighting, and that refusal-based safeguards can be circumvented. The proposed defense injects perturbations into a diffusion model’s latent space while preserving perceived image quality.
Contextual redaction and recommendation with guarantees
ASIRF: An Agentic Framework for Context-Dependent Sensitive Information Redaction
Submitted on 24 September 2026; announced on 25 September 2026.
The framework assumes that information sensitivity depends on domain and purpose: instead of a taxonomy fixed at training time, it retrieves domain-specific definitions from a knowledge base at inference time, without retraining. In the tests, its recall exceeds that of a trained classifier in 68 of 80 model–domain combinations.
Decoupled Learning and Selection in Slate Recommendation for Privacy and Stability Under Noisy Scores
Submitted on 24 September 2026; announced on 25 September 2026.
Slate recommendation is formalized as a randomized scoring model followed by deterministic selection: a properly scoped differential privacy guarantee carries over, through post-processing, to the selection and its audit trace, but only under specific conditions on the selector’s inputs. The contribution is useful for those who need to demonstrate the auditability of recommender systems.
AI ACT IN A NUTSHELL: Part 40
Article 44 - Certificates
In the last installment, we examined Article 43 and the three conformity assessment routes: for biometric systems under Annex III, point 1, the choice between internal control and Annex VII; internal control for systems under Annex III, points 2 to 8; and the sectoral procedure for systems under Annex I, Section A, as rewritten by Regulation (EU) 2026/1744. When a notified body is involved under Annex VII, the assessment results in a certificate. Article 44 governs its language, validity, extension and subsequent developments.
What notified bodies certify
Under the Annex VII procedure, the notified body assesses the quality management system and the technical documentation. If the system complies with the requirements of Chapter III, Section 2, it issues a Union technical documentation assessment certificate, stating the provider, the conclusions of the examination, any conditions of validity and the data needed to identify the system (Annex VII, point 4.6). The same body assesses changes that could affect compliance or the intended purpose and decides whether a new assessment is required under Article 43(4) or whether a supplement to the certificate is sufficient (Annex VII, point 4.7).
Language and validity
Paragraph 1 provides that certificates issued under Annex VII are drawn up in a language that can be easily understood by the relevant authorities of the Member State in which the notified body is established.
Paragraph 2 sets the maximum validity: certificates are valid for the period they indicate, which may not exceed five years for AI systems covered by Annex I and four years for those covered by Annex III. At the provider’s request, validity may be extended for further periods, within the same limits, based on a re-assessment under the applicable procedures. A supplement remains valid as long as the certificate it supplements is valid.
Expiration is therefore not a formality: extension presupposes a re-assessment, which should be planned well ahead of the date indicated on the certificate.
Suspension, withdrawal and restrictions
Paragraph 3 covers the case where the notified body finds that the system no longer meets the requirements of Section 2. Taking account of the principle of proportionality, the body suspends or withdraws the certificate or imposes restrictions on it, unless the provider ensures compliance through appropriate corrective action within an appropriate deadline set by the body. The decision must state reasons.
The provision does not establish an automatic sequence of tolerance: the measures do not apply if the provider ensures compliance through appropriate corrective action within the deadline set by the notified body, and the choice of measure remains governed by the principle of proportionality. The refusal, restriction, suspension or withdrawal of certificates is also subject to the information obligations of Article 45, which we will examine in the next installment.
Appeal
The last subparagraph of paragraph 3 requires an appeal procedure against decisions of notified bodies, including decisions on conformity certificates issued. This safeguards the provider: a decision affecting the placing of the system on the market must be open to review.
Article 44 and Regulation (EU) 2026/1744
Regulation (EU) 2026/1744 amends neither Article 44 nor Annex VII. It does, however, change the context: new Article 43(3) refers, for systems under Annex I, Section A, to points 3, 4.3, 4.4 and 4.5, the fifth subparagraph of point 4.6, and point 5 of Annex VII; Articles 28 and 29 introduce the single application and unified assessment procedure for bodies already notified under the harmonization legislation in Annex I, Section A. For products in that Section, it is therefore necessary to check case by case how the certificates required by sectoral legislation relate to the AI Act requirements.
Practical implications
For providers, DPOs and compliance officers, three operational points.
First, record the certificate’s expiration date in the quality management system and plan the re-assessment needed for extension, within the four-year limit for Annex III systems and the five-year limit for Annex I systems.
Second, manage changes in coordination with the notified body. The distinction between a substantial modification, which requires a new assessment, and a change that can be handled through a supplement depends on the technical documentation and on the pre-determined change envelope (Article 43(4)).
Third, put procedures in place to respond to a finding of non-compliance within the deadline set by the body, and know in advance which appeal procedure is available.
On timing, Article 44 belongs to Chapter III, Section 5, which is not covered by the postponement that Article 113 provides for Sections 1, 2 and 3; however, the Section 2 requirements that certificates attest apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems. Certificates must therefore be read together with the application dates of the requirements they attest.
Legislative sources:
|
Next installment
In Part 41, we turn to Article 45 - Information obligations of notified bodies, which governs notified bodies’ communications to the notifying authority and to other notified bodies: certificates and approvals issued, refused, suspended, or withdrawn; requests from market surveillance authorities; and the confidentiality of information.
LEGAL PROMPTING
The published installments remain available in the newsletter archive: a guide to the professional use of language models in legal practice, from prompt structure to verification of sources and outputs.
Legal Prompting: the new frontier of AI in the legal field
PODCAST
NicFab Podcast episodes, including those on Legal Prompting, remain available. We’ll announce any new series through these channels.
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 standardization 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 organization 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 afterward. 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 characterization, 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
Legislative Decree 160/2026: biometrics, documentation and liability in the implementation of Italian Law 132/2025
28 September 2026
Italian Legislative Decree 160/2026: AI in policing, facial recognition, impact assessments, Article 437-bis of the Criminal Code and Legislative Decree 231/2001, access to evidence and presumption of causation in damage claims. An overview of the first decree implementing Law 132/2025.
Featured events and meetings
Events after the review period; dates below refer to the events, not the publication of their announcements.
Rencontres Informatique & Libertés (scheduled: 29 September 2026)
CNIL |
Stakeholder event on guidelines on the interplay between data protection and competition law (scheduled: 15 October 2026)
EDPB |
Digital Talent EU Days 2026 in Dublin (scheduled: 15-16 October 2026)
European Commission |
Pitching day: Finalists of the Apply AI Startup Award (scheduled: 20 October 2026)
European Commission |
Inaugural D-TECT Forum — Drone Tech for Countering Threats (scheduled: 11 November 2026)
European Commission |
Apply AI Summit (scheduled: 17 November 2026)
European Commission |
Conclusion
In my view, this week shows data protection enforcement paying increasing attention to method. The EDPB’s Guidelines 04/2026, open for consultation until 13 November, propose a five-step approach to choosing between a fine and other corrective powers; IMY’s request for an opinion on Article 10 GDPR likewise seeks a uniform interpretation. The DPC’s decision on Google and the AEPD’s decision on Securitas Direct, although the latter dates from 2023, show how much transparency, accountability and the practical exercise of rights weigh.
On security, ENISA finds that, among unauthorized-access incidents with an identified intrusion vector — only 5% of cases — 60% exploited a vulnerability. The figure should be read with that caveat. CERT-EU’s advisory on the actively exploited vulnerability in F5 BIG-IP APM illustrates this risk. IMY’s fine against Miljödata brings Article 32 GDPR back to its most concrete dimension: controls on software installation and intrusion monitoring.
The theme of human control runs through different contexts. The declaration promoted by Finland and Norway, also signed by the Commission, calls for pre-deployment testing and independent evaluation of frontier models, but it remains a statement of political direction. In the world of work, the Garante President’s interview highlights the limits on automated decisions affecting platform workers’ conditions.
Finally, the implementing regulations on the European Health Data Space and the good practices under Article 18 DSA differ in nature but serve a similar purpose: translating general obligations into verifiable procedures. In my view, the ability to demonstrate compliance in practice depends on that translation.
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.