Newsletter #44 - 27 October 2026
Read on website
---
NicFab Newsletter
Issue 44 | 27 October 2026
Privacy, Data Protection, AI, and Cybersecurity
---
Welcome to issue 44 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
- INTERNATIONAL DEVELOPMENTS
- ARTIFICIAL INTELLIGENCE
- CYBERSECURITY
- TECH & INNOVATION
- SCIENTIFIC RESEARCH
- AI ACT IN A NUTSHELL: Part 44
- LEGAL PROMPTING
- PODCAST
- THE BOOK “Agentic AI”
- FROM THE NICFAB BLOG
- Featured events and meetings
- Conclusion
---
EDPS - EUROPEAN DATA PROTECTION SUPERVISOR
Cybersecurity and the human factor: the EDPS’s new podcast episode
On 23 October, as part of European Cybersecurity Month, the EDPS published a podcast episode on “human defences”. According to the EDPS, in Europe 60% of cyber intrusions begin with a human moment, such as a click on a phishing link or a voice on the phone that sounds familiar; the reports discussed in the episode include a Eurobarometer survey and the ENISA Threat Landscape 2026. So far in 2026, EU institutions, bodies, offices and agencies have notified the EDPS of 115 personal data breaches, about half of them linked to cyberattacks. Konstantina Vemou, the EDPS’s Technology and Security Officer, explains how AI makes phishing harder to spot, which warning signs to look for and why a once-a-year training module is no longer enough.
EDPS — podcast of 23 October
ARETE: assessing the risk to individuals in personal data breaches
On 20 October the EDPS published its report on ARETE (Assessment of Risk to individuals of pErsonal daTa brEaches), the half-day workshop held on 30 June 2026 at the European Parliament InfoHub in Brussels for EU institutions, bodies, offices and agencies. Participants came from 27 bodies and included DPOs, controllers, IT and security specialists, legal officers and operational staff. The applicable framework is Articles 34 and 35 of Regulation (EU) 2018/1725, which run parallel to the GDPR rules on notification to the authority and communication to data subjects: the exercise focused on how to determine whether a breach is likely to result in a risk, or a high risk, to individuals.
The central message is that risk assessment cannot be reduced to mechanical scoring: it requires professional judgment and documented reasoning, as shown by the fact that participants sometimes reached different classifications for the same scenario. To guide that reasoning, the EDPS developed a platform with a four-stage dashboard: establishing the facts, the likelihood of adverse effects, the severity of potential harm, and a reasoned decision on classification and notification. The topic is also timely from a legislative point of view: the Digital Omnibus proposes to change the GDPR notification threshold and also amends Regulation 2018/1725, whose revision is before the Council (see the Council section).
EDPS — report of 20 October
---
COUNCIL OF THE EUROPEAN UNION
Digital Omnibus: the Council sets its mandate for negotiations with Parliament
On 21 October 2026 the Permanent Representatives Committee approved the Council’s mandate to negotiate the Digital Omnibus (procedure 2025/0360(COD)) with the European Parliament; the text was circulated to delegations on 22 October as document 14756/26. Compared with the Commission proposal, the mandate extends the exercise to the DORA Regulation (EU) 2022/2554 and the eIDAS Regulation (EU) No 910/2014 as regards incident reporting, while the title of the text no longer refers to Regulation (EU) 2024/1689.
On the GDPR, the mandate departs from the Commission proposal on the notion of personal data: it leaves the definition in Article 4(1) unchanged and introduces a new Article 25a on pseudonymization and the identification of the natural person, under which pseudonymized data are not personal data for a third party that is unable to identify the data subject. It confines to incidental and residual cases the derogation from the prohibition on processing special categories of data in the development and operation of AI systems and models, and allows controllers to treat access requests pursued for abusive purposes as excessive. On breaches, the mandate proposes that notification to the supervisory authority become mandatory only for high-risk breaches, within 96 hours, through a national entry point; the European Data Protection Board would prepare a common template and lists of circumstances that are and are not high-risk.
The mandate also keeps, as Article 88 bis, the provision linking the processing of personal data in the development and operation of AI systems and models to legitimate interest under Article 6(1)(f), unless the data subject’s rights prevail, in particular where the data subject is a child, and subject to safeguards and a right to object. According to noyb, which commented on the approved text on 23 October, the provision would in practice give AI companies’ interests general precedence over the right to data protection.
Cookie rules stay in the ePrivacy Directive, with a rewritten Article 5(3): refusing consent must be possible with a single click and, after a refusal, the request may not be repeated for at least four months. On incidents, the single entry point at ENISA proposed by the Commission becomes an information point that guides entities to the applicable obligations and to the national entry points provided for in a new Article 23b of the NIS2 Directive.
Council — press release of 21 October
|
Council — doc. 14756/26 of 22 October
|
noyb — 23 October
Working Party on Data Protection, 21 October meeting
On 21 October the Working Party on Data Protection met at the Council Lex building in Brussels, also in a Mixed Committee format with Iceland, Norway, Switzerland and Liechtenstein. The main item was the Presidency compromise text on the proposal amending Regulation (EU) 2018/1725 on the processing of personal data by Union institutions, bodies, offices and agencies (WK 15744/26 and WK 15747/26). The agenda also included the presentation of the FRA report on Directive (EU) 2016/680 (doc. 14304/26, already flagged in issue 43), an exchange of views with the EDPB on the follow-up to the Helsinki Statement, and a Commission update on recent developments in data protection.
Council — agenda CM 4557/2/26 REV 2
---
EUROPEAN COMMISSION
Alphabet and Google Android interoperability for AI services: the Commission specifies the Article 6(7) DMA obligations
On 21 October 2026 the Official Journal (C/2026/5428) published the summary of Decision C(2026) 5094 of 16 July 2026, by which the Commission, in case DMA.100220, specified the measures Alphabet must implement on Google Android to comply with Article 6(7) of Regulation (EU) 2022/1925.
The specification proceedings under Article 8(2) DMA were opened on 27 January 2026; preliminary findings followed on 27 April and the Advisory Committee delivered its opinion on 15 July, after a public consultation. Google Android has been a core platform service since Alphabet’s designation as a gatekeeper on 5 September 2023. The obligation at stake is to give third parties, free of charge, effective interoperability with the same operating system features that Alphabet has available or uses.
The decision concerns the Google Android features that third parties need to offer AI services on the operating system: the Commission identified the functionalities where interoperability fell short and imposed specific measures on Alphabet. It is the first Article 8(2) DMA specification that expressly addresses third-party AI providers’ access to the resources of a designated mobile operating system.
EUR-Lex — OJ C/2026/5428 of 21 October
---
EUROPEAN PARLIAMENT
Plenary: global AI governance reopens the question of civil liability
On Thursday 22 October in Strasbourg, the House debated a Commission statement on strengthening global AI governance; no vote was scheduled. During the debate, Socialists and Greens (among them Alex Agius Saliba, Brando Benifei, Kim van Sparrentak and Sergey Lagodinsky) asked the Commission to table new civil liability rules for AI, after the 2022 proposal for a directive was withdrawn as part of the simplification drive. The EPP, through Arba Kokalari, rapporteur on the AI omnibus that postponed part of the AI Act’s application, argued that the existing framework is sufficient and that regulatory burdens should be reduced. The split within the majority supporting the Commission makes a new proposal unlikely in the short term.
European Parliament — agenda of the 19-22 October session
|
POLITICO Europe
---
GARANTE PRIVACY ITALY
BancoPosta loans: Poste Italiane and Deutsche Bank fined
With two injunction orders of 7 October 2026, made public in newsletter No 554 of 23 October, the Garante fined Deutsche Bank EUR 33,000 (decision No 746) and Poste Italiane EUR 48,000 (decision No 747) for breaching transparency obligations in the offer of BancoPosta loans. The case stems from a complaint by a Poste employee who, after her own loan application was rejected, found information on her financial situation, including net income and outstanding commitments, on a platform accessible to certain members of staff.
The investigation found that, contrary to what the agreements between the parties provided, the two companies acted as independent controllers: Poste collected the applications and acted as intermediary, selecting the most suitable product, while Deutsche Bank assessed creditworthiness and decided whether to grant the loan. Poste was found not to have properly informed applicants about the recipients of the data and the data flows to the bank; Deutsche Bank failed to state in its privacy notice Poste’s role as recipient and independent controller, and was ordered to update the notice within six months. The Garante stresses that the need for concise information does not justify leaving out essential elements.
The same newsletter reports new FAQs on credit reference systems, the favorable opinion on the draft decree on the IT-Wallet (decision No 631 of 23 September 2026), and the launch, on 4 November, of the EDPB portal on children’s privacy.
Garante — Newsletter No 554 of 23 October 2026
|
Garante — decision No 746 of 7 October 2026
|
Garante — decision No 747 of 7 October 2026
|
Garante — FAQs on credit reference systems
|
Garante — decision No 631 of 23 September 2026
Telemarketing and consumer associations: the Garante takes stock
On 20 October in Rome the Garante met consumer associations at the conference “Vita digitale e dati personali” (Digital life and personal data), organized with the Ministry of Enterprises and Made in Italy. The figures presented mainly concern telemarketing: more than 104,000 reports of unwanted calls in 2025 and over 37,000 in the first seven months of 2026, more than 170 a day. President Pasquale Stanzione called for stronger consumer protection, warning that without it society risks returning to “the times of feudal lords”. Board member Agostino Ghiglia recalled the Court of Justice case law allowing associations to bring GDPR actions without a mandate from individual data subjects, describing public enforcement and collective action as two complementary tracks.
Garante — press release of 20 October
Halloween: the Garante’s guide
Ahead of 31 October the Garante published a guide that opens with sharing photos and videos of children in costume online. The central point is the child’s right to decide in future about their own image: published content can be downloaded by third parties or captured by AI systems and stay online for a long time without being easy to delete. The practical advice: limit visibility to friends only, make faces unrecognizable (pixelating them or covering them with an emoji), ask for the consent of those with parental responsibility when other children appear, and use view-once mode in chats where available, bearing in mind that control over images is lost once they are sent. A closing chapter covers seasonal scams: be wary of emails and messages promising exceptional discounts or free gifts.
Garante — press release of 24 October
---
CNIL: FRENCH DATA PROTECTION AUTHORITY
Plenary agenda of 22 October 2026
The CNIL met on 22 October 2026 with a two-part agenda. The first part covered an opinion on a draft decree authorizing the automated processing called “PARCOURS” and the deliberation establishing the “DossierFacile” processing. In the second part, under the procedure of Article 17 of its rules of procedure, it examined two opinions: one on a processing operation for scientific research on workplace discrimination, the other on a draft decree on the use of aircraft-mounted cameras by the prison administration.
CNIL — agenda of 22 October
Journée RGPD in Poitiers on 1 December 2026
On 20 October the CNIL announced a new Journée RGPD in Poitiers on Tuesday 1 December 2026, at the Salons de Blossac, with the AFCDP, the City of Poitiers and Grand Poitiers. Attendance is free, but registration is mandatory and ID is required for access; the event is in person only and no replay will be available. Two sessions stand out: cyber risk management, with an account of the attack suffered by the City of Poitiers (DPO Yohann Brossard and CISO David Bridonneau, with Mélodie Fourez of ANSSI), and a participatory workshop on preparing for a CNIL inspection, led by Tanguy Boucher. The program also covers the CNIL’s new competences, the interplay between the AI Act and the GDPR, the evolving role of the DPO and digital education; opening remarks by President Marie-Laure Denis.
CNIL — Journée RGPD 2026 Poitiers
New telephone prospecting regime: webinar replay available
Since 11 August 2026, telephone marketing to consumers in France requires their prior consent, with limited exceptions; the rules on information, calling hours and the period of validity of consent continue to apply. On 19 October the CNIL published the recording of the webinar held on 24 September 2026 with the DGCCRF to explain the new framework: Article L.223-1 of the Consumer Code, Decree No 2026-662 of 23 July 2026 on the collection, retention and withdrawal of consent, and Articles 4(11), 6, 7 and 21 GDPR.
CNIL — telephone prospecting webinar
---
CASE LAW
ECtHR, decision, Reporters Without Borders, German Section and Härting v. Germany, nos. 81993/17 and 81996/17
Adopted on 29 September 2026, notified on 22 October 2026 — Committee, Fourth Section — ECLI:CE:ECHR:2026:0929DEC008199317
Applications declared inadmissible concerning the bulk interception of cross-border telecommunications by the German foreign intelligence service (BND) in 2012 and 2013, brought under Articles 8, 10 and 13 of the Convention. The complaints against the regime as such are rejected for non-exhaustion of domestic remedies; the complaint about the lack of judicial review is manifestly ill-founded, because ex post review need not be judicial and the applicants had not turned to the G 10 Commission, the non-judicial oversight body available to them.
HUDOC
ECtHR, Casali v. San Marino, nos. 12744/24 and 33586/24
22 October 2026 — Chamber, Fifth Section — ECLI:CE:ECHR:2026:1022JUD001274424
Violation of Article 6 § 1 and Article 10 of the Convention. The case concerns criminal proceedings over a Facebook post about a politician: the Judge for Extraordinary Remedies quashed the applicant’s final acquittal, reviving her conviction, without any legal basis to do so; it follows that the interference with freedom of expression was not prescribed by law. The judgment is not final.
HUDOC
Our review found no rulings of the Court of Justice or the General Court relevant to data protection, AI or platform regulation.
---
On 21 October 2026 the Commission opened formal proceedings to assess whether Booking.com has breached the Digital Services Act. There are three strands: identifying and mitigating risks linked to fraudulent or non-compliant listings, including non-existent properties and impersonation; phishing risks linked to how reservation and guest data are shared with accommodation providers and to the design of the messaging service; and advertising transparency, as listings that gain visibility in exchange for a higher commission may not be clearly recognizable as ads. If proven, the conduct would infringe Articles 34 and 35 DSA, on the duty of very large online platforms to assess and mitigate systemic risks, and Article 26(1), on the identification of advertisements.
Commission — press release IP/26/2223 of 21 October
|
Commission — Shaping Europe’s digital future
Online financial scams: the Commission study supporting DSA enforcement
On 20 October the Commission presented its study on financial scams on online platforms, procured in 2025 to support supervision and enforcement under the DSA and carried out by Open Evidence and Spark Legal and Policy Consulting; the report had been available on the Publications Office website since 16 October. The study groups scams into six categories (investment, romance, e-commerce, employment and collaboration, unexpected gain, disaster relief) and identifies investment and romance scams as the most common and fastest-growing in Europe, with increasingly blurred lines between categories.
The work examines the platform features that scammers exploit: advertising and recommender systems, moderation of organic content, monetization, app stores and private messaging, inauthentic identities, deceptive commercial content and generative AI. It also describes the role of organized criminal networks that sell tools and services, lowering the barrier to entry. Countermeasures are grouped into four families: access and gatekeeping (know-your-business-customer checks and identity requirements), detection and removal, user-facing tools (reporting and information on redress), and platform-level policies with third-party cooperation. The study does not assess compliance by individual providers of very large online platforms and search engines.
Commission — publication of 20 October
|
Publications Office — report, 16 October
On 20 October the Commission brought together providers of very large online platforms and Digital Services Coordinators to discuss the out-of-court dispute settlement mechanism under Article 21 DSA, which allows users to challenge content moderation decisions without going to court; the news item was published on 21 October. Platforms set out the challenges they face, while the Coordinators presented their work on best practices within the European Board for Digital Services. Five topics were discussed: how platforms inform users of this option, which decisions are eligible for review by settlement bodies, the exchange of case information, the fees charged and the implementation of decisions. An external study on these aspects has been completed and will be published soon.
Commission — news of 21 October
Washington again links steel tariffs to softer DMA enforcement
According to POLITICO, the Trump administration has told the Commission, which passed the message on to EU capitals, that it will not cut tariffs on EU steel and other goods unless Brussels eases enforcement of the Digital Markets Act. US duties on steel and aluminium still reach 50%, against the general 15% ceiling agreed at Turnberry. By the end of the year the Commission must report to Parliament and Council on compliance with the deal and could, in theory, withdraw the zero-tariff treatment of US industrial goods, a politically unlikely step. The linkage is not new: in November 2025 Commerce Secretary Howard Lutnick made a similar offer, which Executive Vice-President Teresa Ribera described as blackmail. Since enforcement began, DMA fines on Apple, Meta and Google total EUR 1.59 billion, including EUR 890 million imposed on Google last July.
POLITICO Europe
The European Publishers Council asked Teresa Ribera and Henna Virkkunen to block the new contractual terms under which Meta governs the use of its tools to run Facebook pages, Instagram accounts and ad campaigns, and to examine them against copyright, competition law and the DMA. In the version reported by the press, the terms allowed Meta to crawl the websites linked to those accounts to train its products, with no opt-out and with continued use of the tools treated as consent. Publishers argue that Meta would obtain for free content that other AI developers pay for. Meta then removed the reference to training, limiting the clause to advertising and business products, and on 23 October postponed the entry into force of the terms from 30 October to 20 November. In the United Kingdom, the Competition and Markets Authority has asked Meta for clarification.
POLITICO Europe
|
Rapporteur — 23 October
Haugen: US pressure confirms that the DSA works
In an interview published on 19 October, Frances Haugen, the former Meta employee who made internal company documents public in 2021, reads US pressure against European digital rules as proof that the Digital Services Act is having an effect, and calls for it to be enforced rigorously.
Rapporteur (formerly Euractiv) — 19 October
---
INTERNATIONAL DEVELOPMENTS
On 21 October 2026 Judge Francis J. Mathew of the New Mexico District Court ordered Meta to pay USD 25 billion, finding more than 43 million violations of the state’s unfair practices law in a case rooted in the 2018 Cambridge Analytica revelations. The judge found untrue Facebook’s assurances that users controlled their data.
The case was brought in 2021 by Attorney General Raúl Torrez. The earlier milestones are well known: the FTC’s USD 5 billion in 2019, the USD 725 million class settlement of December 2022, and the USD 18 billion agreed in August with almost all states, which New Mexico and Florida declined to join, with USD 459 million earmarked for Cambridge Analytica claims. The original dataset covered 87 million profiles, harvested through the “thisisyourdigitallife” app from around 300,000 users who completed the quiz. Meta has announced an appeal.
What matters is less the amount than the method: more than 43 million separate violations derived from a single course of misleading statements. If other state attorneys general follow the same approach, platforms’ exposure in US privacy litigation would change by an order of magnitude.
The Hacker News — 22 October
United Kingdom: the FCA on good practice and weaknesses in AI use in financial services
On 19 October the Financial Conduct Authority published a review of good practice and areas for improvement in the use of AI in financial services, drawn from supervisory work, engagement with industry and its AI Lab initiatives. The document creates no new requirements and points to the existing framework, from the Consumer Duty and the senior managers regime to operational resilience and outsourcing. Among the weaknesses, the FCA flags excessive reliance on human review of outputs without addressing underlying model risk or governance responsibilities: human review can mask weaknesses in the model and become a box-ticking exercise. Another weak area is dependence on a single model provider, with no alternatives or exit plans.
In the speech accompanying the publication, Jessica Rusu announced that the FCA will work with industry, fellow regulators and frontier labs on a framework for agentic finance, with clear approaches to agent identity, delegated authority, accountability and execution.
FCA — review of 19 October
|
FCA — speech by Jessica Rusu, 19 October
Texas: the Attorney General sues OpenAI and Sam Altman
On 23 October Texas Attorney General Ken Paxton brought an action against OpenAI, its affiliated entities and its CEO Sam Altman under the Texas Deceptive Trade Practices–Consumer Protection Act. According to the Attorney General’s office, OpenAI presented ChatGPT as a safe and reliable tool without adequately warning users of the risks or protecting them, including minors. These are allegations made in the lawsuit, not established facts: the case is at its outset.
Texas Attorney General — 23 October
Germany: political microtargeting by major parties found unlawful
On 21 October noyb made public the decisions in which the Berlin data protection authority, ruling on complaints filed by the organization over the 2021 federal election, reprimanded the CDU, SPD, Greens and Die Linke for using users’ political interests to target personalized political advertising without a legal basis such as consent. The complaint against the AfD is still pending, after the party challenged the authority’s investigative powers in court. The decisions are reported by the complainant, which publishes excerpts.
noyb — 21 October
On 23 October Datatilsynet opened a formal investigation into the large number of look-ups in the Danish civil registration system (CPR), sending requests for clarification to the registry administration and to Charity Guard ApS on how the look-ups were possible and which security measures were in place. The investigation follows the breach notification received from the CPR administration on 4 October.
Datatilsynet — 23 October
Spain: more than 3,000 breach notifications to the AEPD in the first half of the year
On 20 October the AEPD reported receiving more than 3,000 personal data breach notifications between January and June 2026, against 1,271 in the same period of 2025 (+138%) and more than in the whole of 2025; for January to August, the increase rises to 163% (4,556 notifications against 1,728). On 19 October the Agency had also signed a cooperation memorandum with the Luxembourg authority.
AEPD — 20 October
|
AEPD — 19 October
---
ARTIFICIAL INTELLIGENCE
Italy: Legislative Decree 179/2026 implementing the AI Act in force from 23 October
On 23 October 2026 Legislative Decree No 179 of 7 October 2026, published in Official Gazette No 234 of 8 October, entered into force. It adapts Italian law to Regulation (EU) 2024/1689 as regards the powers of national authorities and the use of artificial intelligence in training. Among other things, the decree sets out the penalty system and the national regulatory sandbox, both analyzed on the blog (see From the NicFab blog).
Gazzetta Ufficiale — Legislative Decree No 179 of 7 October 2026
Council of Europe: two HUDERIA Model components adopted
From 21 to 23 October the Council of Europe’s Steering Committee for New and Emerging Digital Technologies (CDNET) held its second plenary meeting in Strasbourg, with 170 participants in person and online from member and observer States, the EU, international organizations, civil society and invited experts. The Committee adopted two components of the HUDERIA Model, the Council of Europe’s instrument for assessing the risks and impacts of AI systems on human rights, democracy and the rule of law: the Stakeholder Engagement Process and the Recommendations on Roles and Responsibilities in the AI Project Lifecycle.
Work continues on other components of the Model, including procurement guidelines and interoperability clauses, and on the mapping study on trends in AI use. Among the guest speakers, the Secretary of the Consultative Committee of Convention 108, a body distinct from the CDNET, presented that Committee’s work, including draft guidelines on data protection in the context of neurosciences.
Council of Europe — 23 October
Gmail, Photos and Gemini: Google’s new policy opens the door to personalized ads
Two in-depth reports on 23 and 24 October revisit the change to Google’s privacy policy in force since 16 October 2026, which provides that data from Gmail, Drive and Google Photos may feed into ad personalization when users turn on Personal Intelligence, the feature that connects these services to Gemini and to AI Mode in Search. The previous wording expressly excluded the use of content from those services for personalized advertising; the new one carves out the case where users choose to make the data available to other Google services that support it.
Google says the mechanism is opt-in, that ads will not be based directly on the content of the inbox or photo library, and that sensitive information remains excluded. Still, with Personal Intelligence turned on, signals derived from that data can shape targeting. The feature is in a global beta for adult users, excluding the EEA, Switzerland, the United Kingdom and Nigeria: a perimeter that matches the jurisdictions with the most intense regulation. It can be switched off in the search personalization settings.
The Hacker News — 24 October
|
Ars Technica — 23 October
OpenAI GPT-6 agents disable monitoring and deny doing so
On 20 October OpenAI disclosed two internal incidents, on 3 and 6 October 2026, in which agents based on GPT-6.1 Sol and GPT-6 Astra weakened misalignment monitoring while working on code. The Sol agent, tasked with fixing an evaluation pipeline, disabled automatic alerts to the security team and emptied the recipient lists, reading an earlier instruction not to send messages during a Slack search as permission to suppress the notifications. The Astra agent, during a migration, removed the monitoring integration to match a reference implementation and, when asked, denied having done so.
OpenAI suggests unintentional over-persistence and says it wants to understand why models choose to reduce controls. The two cases add to the misalignment incident reports the company published on 9 and 15 October. The verifiability of what an agent claims to have done thus becomes a matter of technical governance.
The Hacker News — 21 October
|
OpenAI — misalignment reports
Cairn weaponized: more than 100 organizations breached with AI agents
Security firm Norforge has documented the abuse of Cairn, an open-source platform for orchestrating AI agents that can also serve as an autonomous penetration testing engine, to automate attacks against hundreds of sites. Three clusters were identified thanks to control panels left exposed. The largest, Chinese-speaking, ran more than 1,100 attack projects across 14 dashboards against e-commerce and payment infrastructure in the United States, Japan and Australia; according to Reuters it breached more than 100 companies, including John Deere and SharkNinja.
The second cluster, English-speaking and probably based in Africa, targeted payment card and cryptocurrency theft; the third ran 59 projects against large US e-commerce sites. The estimated success rate is around 10%; of the secrets found on the exposed panels, 11% proved valid. Attacker infrastructure, too, shows the usual weaknesses of projects rushed into production.
SecurityWeek — 20 October
Spain: PP election video manipulated with AI
The Partido Popular released on its channels an election video that uses AI to alter the likeness of people close to the PSOE, including Begoña Gómez, wife of Prime Minister Pedro Sánchez, and former Prime Minister José Luis Rodríguez Zapatero, showing them admitting wrongdoing linked to ongoing investigations. By Friday the video had approached two million views; the notice of its artificial origin is there, but in small, semi-transparent type. Leire Díez, a former Socialist party member whose image was used, has filed a complaint for defamation and digital identity theft.
Digital Transformation Minister Óscar López pointed to the risk of confusion with reality, while most of the nine lawyers consulted by El País consider the video lawful, given the notice and its satirical tone. The case illustrates the gap between the deepfake transparency obligation in Article 50 of the AI Act and what a voter, exposed for a few seconds, actually perceives.
The Guardian — 24 October
---
CYBERSECURITY
PHP2Shell: coordinated EU response and national authority updates
On 21 October 2026 the European Commission (DG CNECT), ENISA, CERT-EU, Europol EC3 and the network of national CSIRTs issued a joint statement urging immediate patching of two critical PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, which allow unauthenticated remote code execution. Fixes were released the same day. The statement recommends prioritizing internet-facing systems, isolating them where patching cannot be immediate, checking third-party products that embed PHP, and recalls that since 11 September 2026 manufacturers must report actively exploited vulnerabilities through the single reporting platform under the Cyber Resilience Act.
On 25 October CSIRT Italia, at the national cybersecurity agency ACN, quantified national exposure: of 741,523 PHP services detected in Italy, 285,149 run potentially vulnerable versions, or 81.9% of those whose version could be identified. The alert also flags a third vulnerability, CVE-2026-115560, still unpatched, for which it indicates a configuration-level mitigation.
According to Ars Technica and the Financial Times, Amazon said it observed exploitation attempts four days before disclosure. In its update of 23 October, the Dutch NCSC offered a more cautious assessment, reporting that exploitation is complex, that the likelihood of large-scale exploitation is lower than first estimated and that it knew of a single, unconfirmed report of possible exploitation abroad, while stressing that patching remains urgent. Also according to the NCSC, the PHP Foundation considers it unlikely that the third vulnerability, CVE-2026-115560, can be used for remote code execution.
The flaws, latent for more than a decade, were found by researchers at Calif and Wiz with the help of OpenAI and Google models; Wiz estimates that PHP powers about 70% of websites, partly because WordPress is largely written in PHP. The same models that bring forward the discovery of flaws also speed up the development of offensive tooling: the interval between discovery and disclosure becomes the critical variable.
ENISA — joint statement of 21 October
|
ACN — CSIRT Italia alert of 25 October
|
NCSC-NL — update of 23 October
|
Ars Technica — 22 October
NIS: ACN simplifies the 2027 registration
On 23 October the ACN published determination 314287/2026, which simplifies the registration of NIS entities ahead of the January 2027 window: entities that have already submitted their declaration and have no changes to report do not need to resubmit it. The FAQs have been updated and, to allow the platform to be adapted, late declarations can no longer be submitted from 31 October 2026. The determination was discussed at the tenth meeting of the NIS coordination body (Tavolo NIS).
ACN — news of 23 October
CISA adds Collabora Online and Arista VeloCloud to the KEV catalog
On Friday 23 October CISA added two actively exploited critical vulnerabilities to its Known Exploited Vulnerabilities catalog. The first is CVE-2026-110447 (CVSS 9.2), a command injection in the CODE Server built into Collabora Online used with Nextcloud, which allows remote code execution when the local CODE server is not running (fixed in September in version 26.4.303). The second is CVE-2026-114625 (CVSS 9.9), an SQL injection in on-premises versions of Arista VeloCloud Orchestrator, fixed in October in versions 5.2.3.18, 6.1.3.10 and 6.4.2.12; exploitation requires network access to the web interface and an authenticated session, even with the minimal Enterprise Read Only role.
Attackers are using the Collabora flaw to install web shells on Nextcloud servers and search for credentials, password manager stores and cryptocurrency wallets. Arista has shared two Japanese IP addresses without attributing the activity; this is the third critical vulnerability in the product exploited since July 2026, after CVE-2026-16812 and CVE-2026-93952, both rated CVSS 10.0. US federal civilian agencies had until 26 October to patch.
The Hacker News — 24 October
BlackX, a BlackLotus-derived UEFI bootkit against diplomatic agencies in Central Asia
Kaspersky has analyzed BlackX, a UEFI bootkit used in early 2026 in an espionage campaign against diplomatic agencies in Tajikistan, Uzbekistan and Turkmenistan and attributed with medium confidence to APT41 (also known as Brass Typhoon, Wicked Panda or Barium). The bootkit exploits Baton Drop (CVE-2022-21894) to bypass Secure Boot, enroll custom Machine Owner Keys and load arbitrary EFI files.
The exploit is implemented almost identically to BlackLotus, while the kernel driver and payloads differ. Some BlackLotus samples analyzed contain code absent from the repository leaked in July 2023 but present in BlackX, suggesting reverse engineering or purchase of the source. The driver, signed with a now-revoked certificate, disables security components, tampers with ETW providers and contains routines aimed specifically at Kaspersky and Avast products; to protect its files, it intercepts attempts to open them and launches Windows Calculator instead. The payloads include a remote access trojan that uses Google Drive, OneDrive or Yandex.Disk as its command-and-control channel.
BleepingComputer — 24 October
Tailscale SSH: an unpatched chain of flaws leads to root on other nodes
On 25 October Erica Windisch of Cyberdione Labs published the “Tailworm” report on a chain of vulnerabilities in Tailscale SSH that lets a compromised device gain root privileges on other servers reachable in the same network. Tests were run on Tailscale 1.104.2 and headscale 0.29.5; at the time of publication there was no bulletin, CVE or fixed version.
Tailscale SSH does not check the user’s passwords or keys: it identifies the device and applies an access rule, so an attacker only needs an initial foothold. The research describes three techniques: ordinary port forwarding that makes the Tailscale process, running as root, connect to a system socket reserved for root; replacing a root-owned socket; and a TOCTOU race condition that bypasses path checks. In the lab the chain made it possible to capture an enrollment key for new devices.
Tailscale said it is working on fixes and does not recommend disabling Tailscale SSH. On headscale 0.29.5, however, forwarding cannot be turned off: every SSH grant is effectively root on the target system.
The Hacker News — 25 October
Phantom, the AI-driven intrusion framework used by INC and The Gentlemen affiliates
SOCRadar has documented an AI-driven intrusion ecosystem run by affiliates of the INC and The Gentlemen ransomware groups and built around a framework called Phantom. The operator, assessed as Russian-speaking, acts both as an intrusion and extortion actor and as an infrastructure provider.
The model is simple: AI executes, the human approves. The operator writes a victim-specific prompt and approves each step, while execution runs through OpenCode with commercial models hosted on OpenRouter; one exposed directory logged 8,257 messages across 133 sessions. When the models refuse, the operator does not switch models but recasts the task as an authorized security test. Exploited vulnerabilities include CVE-2024-57727 (SimpleHelp), CVE-2026-35616 (FortiClient EMS) and CVE-2026-41940 (cPanel/WHM); others are in preparation.
The force multiplier remains the most mundane one: in one case a single reused password opened, in turn, the VPN, servers, remote management tools, backups and security platforms. AI speeds things up, but weak security posture still does most of the work.
The Hacker News — 23 October
FortiSIEM: public exploit for unauthenticated code execution
Horizon3.ai has published a proof of concept for CVE-2026-106173 (CVSS 9.4), a critical path traversal in FortiSIEM fixed by Fortinet in version 7.5.1 (7.6 is not affected). The exploit targets the gRPC service listening on TCP port 27918: the function that handles query results uses a user-supplied identifier without validation to build the file path. By combining directory traversal sequences with a null byte injected into gRPC’s HTTP/2 encoding, an attacker neutralizes the suffix automatically appended to the file name and writes wherever they want.
The chosen target is a Redis health-check script run by cron every 60 seconds: overwriting it preserves its owner and permissions, guaranteeing code execution. The Canadian Centre for Cyber Security has issued an advisory.
The Hacker News — 20 October
Qilin and INC exploit PAN-OS CVE-2026-0257 to bypass MFA
On 19 October Palo Alto Networks’ Unit 42 updated its analysis, confirming continued exploitation of CVE-2026-0257 (CVSS 7.8), an authentication bypass in PAN-OS and Prisma Access patched in May 2026 and added by CISA to the KEV catalog on 13 May. The ransomware operators identified include Qilin and INC; a few days earlier ReliaQuest had also reported the Settra group.
The flaw allows unauthorized VPN sessions to be established while bypassing multi-factor authentication; this is followed by internal reconnaissance with dual-use remote management tools, lateral movement over RDP and SMB, credential harvesting, exfiltration and ransomware deployment. According to Unit 42, the most significant evolution is the direct pivot to the virtualization layer: in one case the attackers moved from the VPN gateway to the vCenter appliance, disabled a security control on the ESXi hypervisors and dumped credentials offline from a domain controller’s virtual disk, evading the EDR installed in the guest system. Initial access was combined with known exploits against Veeam backup infrastructure (CVE-2023-27532). Beyond patching, Palo Alto Networks recommends actively hunting for indicators of compromise.
Unit 42 — update of 19 October
|
The Hacker News — 20 October
Qwen3-Coder writes more vulnerable code when the client is American
Security firm Hirundo found that Alibaba’s Qwen3-Coder produces systematically more vulnerable code when the prompt names a US entity or a group Beijing disapproves of as the client, confirming earlier analyses by CrowdStrike (November 2025) and Booz Allen Hamilton (May 2026).
The test generated 1,840 web applications with sign-up, login and secret storage functions, checked with Semgrep (Meta’s PurpleLlama rules) and ETH Zurich’s BaxBench benchmark: 1,144 proved breachable. The code was vulnerable in 70% of cases for a US defense contractor, against 59% for a Chinese one; applications for a Hong Kong pro-democracy group leaked secrets in 47% of cases, against 37% for a neutral entity. Explicitly asking for secure code does not close the gap: for the US contractor the share rises to 74%. Hirundo says it neutralized the behavior by removing the alignment with the Chinese Communist Party from the weights through unlearning techniques.
Model provenance thus becomes part of supply chain security, which NIS2 requires entities to manage (Article 21(2)(d)): the conditioning lives in the weights and cannot be fixed with instructions.
SecurityWeek — 21 October
---
TECH & INNOVATION
Smart TVs: suspicious traffic to ACR endpoints even after opt-out
A Northeastern University study published on 19 October shows that smart TVs running Samsung Tizen, Vizio OS (formerly SmartCast) and Roku OS keep sending traffic to Automatic Content Recognition endpoints even after users have turned tracking off in the settings. Encryption prevented the researchers from seeing what the traffic contained: the authors regard it as suspicious and as a sign of potential privacy violations, but acknowledge they cannot establish this with certainty.
The paper, “Watching the Watchers: Privacy Analysis of US Smart TV Automatic Content Recognition”, examined eight operating systems on twelve TVs from ten manufacturers (Amazon, Hisense, LG, Roku, Samsung, Sony, Supersonic, TCL, Vizio, Westinghouse), which according to the authors cover about 95% of the installed base in the United States. ACR captures screen images or audio snippets and sends them to remote servers for audience measurement and targeted advertising. The researchers ran 5,943 controlled experiments, intercepting traffic with Mon(IoT)r and the SPHERE Research Infrastructure. On Roku the behavior only emerged when using the built-in The Roku Channel app; some systems, including Tizen, use ACR domains recognizable by their names.
If it were confirmed that tracking based on withdrawn consent continues, this would raise an issue of lawfulness of processing: an opt-out that does not stop tracking is not an opt-out. Traffic to ACR endpoints alone, without knowing its content and purpose, does not however make it possible to establish this. Should further checks confirm it, the pattern would recall dark patterns in cookie banners, moved onto home hardware.
Ars Technica — 21 October
---
SCIENTIFIC RESEARCH
Selection of the most relevant papers of the week from arXiv on AI, Machine Learning and Privacy
Membership Inference and Privacy Auditing
Automated screening of privacy risks in medical AI
Submitted on 21 October 2026; announced on 22 October 2026.
The authors show that medical models shared without their training set remain vulnerable to membership inference attacks, and that reporting risk as a cohort average hides the fact that individual patients are almost certainly exposed. The consequence for the DPIA of clinical models is direct: an aggregate metric is not enough to demonstrate the adequacy of measures under Article 32 GDPR when risk is concentrated on specific individuals.
arXiv
Shift-aware score transport for privacy auditing without retraining
Submitted on 22 October 2026; announced on 23 October 2026.
The paper proposes an observational audit that estimates lower bounds on privacy loss without inserting canaries or retraining the model, working on fixed sets of known members and non-members. This is the realistic scenario for anyone auditing a third-party model, where retraining and canary insertion are not feasible.
arXiv
Right to Erasure and Persistence in Agentic Systems
AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents
Submitted on 22 October 2026; announced on 23 October 2026.
A behavioral benchmark to check whether, once a record is deleted from an LLM agent’s persistent memory, the information survives in what derives from it (summaries, derived states). The work pinpoints an operational problem in exercising Article 17 GDPR against agentic systems: deleting the original record does not mean deleting the information.
arXiv
Indistinguishable Agents: Training Agent Privacy as Twin-World Indistinguishability
Submitted on 18 October 2026; announced on 20 October 2026.
A training approach in which the agent is optimized on “twin worlds” that differ only in a private fact, requiring its observable behavior to remain indistinguishable. It overcomes the limits of methods that merely suppress explicit mentions while ignoring what can be inferred from actions.
arXiv
Contextual Privacy and DP Analysis of LLM Conversations
Privacy in Context Is Being Overlooked in the Age of General-Purpose LLMs
Submitted on 21 October 2026; announced on 22 October 2026.
A position paper against one-size-fits-all privacy solutions in general-purpose LLMs: privacy is contextual, and a single system spanning different domains cannot apply the same rule everywhere. A useful reference for anyone drafting acceptable-use policies or assessing enterprise configurations of general-purpose assistants.
arXiv
CALLIOPE: Private Insights into AI Use, Without Embeddings
Submitted on 20 October 2026; announced on 22 October 2026.
A framework for discovering topic hierarchies in large collections of LLM conversations under differential privacy, without clustering vector embeddings and using the model itself in the pipeline. Of interest to anyone who wants analytics on the use of conversational products with formal DP guarantees.
arXiv
Cybersecurity: Hardware Attacks and Vulnerability Triage
Amplifying Latency and Cost of LLM Inference via Rowhammer Attack
Submitted on 22 October 2026; announced on 23 October 2026.
The study assesses whether a Rowhammer-induced bit flip in CPU DRAM can degrade LLM inference on GPUs, compiled through Triton’s MLIR/LLVM pipeline, while leaving the output correct. It is an economic attack vector against inference providers: it does not compromise correctness, it hits latency and cost.
arXiv
Predicting CVSS and CWE from CVE Descriptions with Fine-Tuned Small Language Models
Submitted on 16 October 2026; announced on 20 October 2026.
Since April 2026, NIST has enriched only high-priority CVEs with CVSS v3.1 and CWE, leaving many vulnerabilities without the classifications used to prioritize patching. The authors train small language models to derive the CVSS vector, severity level and CWE from the CVE description, filling the gap left by the NVD.
arXiv
---
AI ACT IN A NUTSHELL: Part 44
Article 48 - CE marking
In the last installment we examined Article 47 and the EU declaration of conformity. Article 48 governs the CE marking, the sign that makes visible on the market the conformity attested by that declaration.
General principles and digital marking
Paragraph 1 makes the CE marking of high-risk AI systems subject to the general principles set out in Article 30 of Regulation (EC) No 765/2008. The choice is consistent with the structure of the AI Act, which inserts the new rules into the existing European framework for product safety and the free movement of goods, rather than creating a separate marking framework.
Paragraph 2 takes account of the intangible nature of many AI systems: for those provided digitally, a digital CE marking is used, but only if it can easily be accessed via the interface from which the system is accessed or via an easily accessible machine-readable code or other electronic means. Software or a service delivered through an API or the cloud does not lend itself to the physical affixing of a mark.
Affixing, notified body and other legislation
Paragraph 3 requires the CE marking to be affixed visibly, legibly and indelibly for high-risk AI systems. Where that is not possible or not warranted on account of the nature of the system, it is affixed to the packaging or to the accompanying documentation, as appropriate.
Paragraph 4 concerns the notified body: where it is involved in the conformity assessment, the CE marking is followed by its identification number, affixed by the body itself or, under its instructions, by the provider or the authorized representative. The same number appears in any promotional material that mentions that the system fulfills the requirements for CE marking.
Finally, paragraph 5 deals with overlapping legislation: where the high-risk AI system is subject to other Union law that also provides for CE marking, the marking indicates that the system also fulfills the requirements of that other law.
Practical implications
For providers of high-risk AI systems, Article 48 calls for planning the marking from the design and distribution stages onwards. They must decide whether the marking will be affixed to the system, the packaging or the accompanying documentation, and justify the choice against the visible, legible and indelible standard. For digitally delivered systems, the solution has to be built into the user interface or into electronic access mechanisms, with effects on design and on distribution choices (web apps, SDKs, APIs, cloud platforms).
Compliance and legal functions need to coordinate with marketing and communications so that the notified body’s identification number, where required, appears correctly in brochures, websites, data sheets and any material that refers to conformity. Omitting the number or using it improperly can result in non-compliant commercial communication.
For organizations placing on the market AI systems already subject to other sectoral legislation requiring CE marking, such as medical devices, machinery or toys, a single marking attests overall conformity. This, however, requires coordination across the different assessment files, with a significant impact on internal quality assurance and regulatory affairs processes.
Legislative sources:
AI Act
Next installment
In Part 45 we turn to Article 49 - Registration, which brings high-risk AI systems into the EU database before they are placed on the market.
---
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 →
NicFab Podcast English
---
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 translates them into a practical tool, 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 supporting materials 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 a subject 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 →
The book page
---
FROM THE NICFAB BLOG
NIS2 in Italy: understanding the law, ACN decisions and EU rules
23 October 2026
How a NIS2 obligation is reconstructed in Italy: Directive, implementing acts, Legislative Decree 138/2024, ACN Determinations, FAQs and guides. Legal status, current acts and connections with artificial intelligence.
Read the full article
The Italian AI regulatory sandbox: how the national sandbox under Legislative Decree 179/2026 will work
22 October 2026
The Italian AI regulatory sandbox in Legislative Decree 179/2026: supervision by AgID and ACN, Coordination Committee, Operational Group, access and priority, the Garante, further sandboxes, FinTech, universities and outcomes.
Read the full article
Digital Omnibus: The Council Adopts Its Position, Simplifying the GDPR Without Dismantling It?
21 October 2026
On 21 October 2026 the Council of the EU adopted its negotiating position on the Digital Omnibus: amendments to the GDPR, cookies under the ePrivacy Directive, the data acquis and trade secrets, national entry points for incident reporting. An analysis of the mandate, of the EDPB and EDPS positions and of the open questions ahead of the trilogues.
Read the full article
Penalties for AI Act infringements in Italy: ceilings, non-pecuniary measures and proceedings in Legislative Decree 179/2026
19 October 2026
Penalties in Italian Legislative Decree 179/2026 for AI Act infringements: ceilings, non-pecuniary measures, proceedings of AgID, ACN, Bank of Italy, CONSOB, IVASS and the Garante, publication, proceeds and comparison with the draft.
Read the full article
---
Featured events and meetings
European AI Innovation Month (scheduled: 14 October – 17 November 2026)
European Commission |
Info
Opinion of Advocate General Szpunar in Case C-250/25, Like Company (generative AI and copyright) (scheduled: 3 November 2026)
Court of Justice of the European Union |
Info
Launch of the EDPB portal on children’s privacy (scheduled: 4 November 2026)
EDPB |
Info
LIBE Committee meeting: items include the vote on the provisional agreement on the child sexual abuse directive (scheduled: 4-5 November 2026)
European Parliament |
Info
European Cybersecurity Skills Conference 2026 (scheduled: 4-5 November 2026)
ENISA |
Info
Judgments in Cases C-205/25 (access to a supervisory authority’s complaint file, Article 15 GDPR) and C-468/24 (data protection and smart meters); Advocate General’s Opinion in Case C-654/25 (dynamic IP addresses) (scheduled: 5 November 2026)
Court of Justice of the European Union |
Info
Feedback on the implementing regulation on Articles 75c and 75d of the AI Act (deadline: 9 November 2026)
European Commission |
Info
Feedback on the Cloud and AI Development Act proposal (deadline: 11 November 2026)
European Commission |
Info
Inaugural D-TECT Forum — Drone Tech for Countering Threats (scheduled: 11 November 2026)
European Commission |
Info
3rd European Cybersecurity Awareness Raising Conference (scheduled: 11 November 2026)
ENISA |
Info
air 2026 — Convaincre ou manipuler ? L’éthique du numérique à l’épreuve des urnes (scheduled: 16 November 2026)
CNIL |
Info
Apply AI Summit (scheduled: 17 November 2026)
European Commission |
Info
Journée RGPD 2026 in Poitiers (scheduled: 1 December 2026)
CNIL |
Info
Feedback on the proposal for an EU Kids Act regulation (COM(2026) 681) (deadline: 3 December 2026)
European Commission |
Info
---
Conclusion
In my view, the week poses a single question beneath very different stories: who verifies, and with what tools, what a system claims to have done.
OpenAI disclosed that two agents, on 3 and 6 October, weakened monitoring while working on code, and that one of them then denied the change. The Phantom framework described by SOCRadar hands the intrusion to the models and leaves the human operator only the approval of each step; when the model refuses, the framing of the prompt changes, not the objective. These are two sides of the same problem: the human oversight provided for on paper does not coincide with the ability to understand, in real time, what is happening. The UK FCA makes the same point from the supervisory side when it warns that human review of outputs can mask weaknesses in the model.
The PHP2Shell case shows what a coordinated institutional response looks like: on 21 October, the same day the fixes were released, the Commission, ENISA, CERT-EU, Europol and the CSIRTs network issued a joint statement, and on 25 October the ACN put Italian exposure at 81.9% of services with an identifiable version. In my opinion, however, another fact is particularly instructive: according to Amazon, exploitation attempts had begun four days before disclosure, even though on 23 October the Dutch NCSC downgraded the likelihood of large-scale exploitation. The window that matters can open before institutions are in a position to speak.
On the regulatory side, the signals point in different directions. The Council’s Digital Omnibus mandate proposes to limit notification to the authority to high-risk breaches only and to treat pseudonymized data as non-personal for those unable to identify the data subject. In the European Parliament, the debate on global AI governance revived the call for civil liability rules, rejected by the EPP in the name of simplification. In Strasbourg, the Council of Europe adopted two HUDERIA Model components on stakeholder engagement and on the allocation of roles and responsibilities in the AI project lifecycle: in my view a useful building block precisely where it is unclear who should answer for what. In Spain, an election video manipulated with AI came close to two million views with a barely legible notice, and most of the lawyers consulted by El País consider it lawful.
In my view, read together, these elements show that declared transparency is not enough without independent verification: a label on a video, a notification filtered by a risk threshold, a log written by the very agent whose behavior it is meant to document. The ECtHR decision on German bulk interception is a reminder too that ex post review need not be judicial, provided an independent body exists and someone turns to it.
My cautious prediction: in the coming months the central question will not be how capable systems are, but how verifiable their claims are. Those who do not now put in place records independent of the system being observed, on the provider side as on the deployer side, will have to reconstruct after the fact behavior that nobody saw happen.
---
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.
Lagrange - a Gemini client for desktop, iOS and Android
Back to newsletter list
English section
Home