Newsletter #46 - 10 November 2026
Read on website
---
NicFab Newsletter
Issue 46 | 10 November 2026
Privacy, Data Protection, AI, and Cybersecurity
---
Welcome to issue 46 of the weekly newsletter dedicated to privacy, data protection, artificial intelligence, cybersecurity and ethics. Every Tuesday you will find a curated selection of the most relevant news from the previous week, with a focus on European regulatory developments, case law, enforcement and technological innovation.
---
In this issue
- EDPS - EUROPEAN DATA PROTECTION SUPERVISOR
- COUNCIL OF THE EUROPEAN UNION
- EUROPEAN COMMISSION
- EUROPEAN PARLIAMENT
- GARANTE PRIVACY ITALY
- CNIL: FRENCH DATA PROTECTION AUTHORITY
- CASE LAW
- DIGITAL MARKETS & PLATFORM REGULATION
- EUROPEAN STANDARDIZATION
- INTERNATIONAL DEVELOPMENTS
- ARTIFICIAL INTELLIGENCE
- CYBERSECURITY
- TECH & INNOVATION
- SCIENTIFIC RESEARCH
- AI ACT IN A NUTSHELL: Part 46
- LEGAL PROMPTING
- PODCAST
- THE BOOK “Agentic AI”
- FROM THE NICFAB BLOG
- Featured events and meetings
- Conclusion
---
EDPS - EUROPEAN DATA PROTECTION SUPERVISOR
On 6 November, the EDPS published episode 23 of its Newsletter Digest, the audio format summarising the previous month’s activity. The episode looks back at the three opinions issued in October on the package with which the Commission intends to reform the EU agencies in the field of justice and home affairs: the changes concern Europol, Eurojust and the data protection rules that apply to these bodies. It then turns to the safeguards in the EU-Republic of Korea agreement on the exchange of passenger name record (PNR) data, the recommendations on cross-border health data and European Cybersecurity Month. As part of the same reform package, the European Parliament published an EPRS briefing during the week, reported below.
EDPS — Newsletter Digest, episode 23 of 6 November
---
COUNCIL OF THE EUROPEAN UNION
EU KIDS Act: the progress report
In a note dated 6 November 2026 (doc. 15088/26), the Presidency sent Coreper and the Council its progress report on the proposed EU KIDS Act regulation (“EU Keeping Internet Digital Spaces Accountable and Trustworthy”, procedure 2026/0286(COD)), presented by the Commission on 17 September 2026. The Audiovisual and Media Working Party has held eight meetings on the file since 6 October and has completed a first examination of Chapters I to VI. The first reading of the text should be completed in early December, when the Irish Presidency plans to table a compromise text on Chapter III. There is no political orientation yet: the document captures a technical negotiation in full swing.
Council — doc. 15088/26 of 6 November
Cloud and AI Development Act: the opinion of the Italian Chamber of Deputies
Document 15470/26 of 5 November forwards to the President of the Council the opinion of the Italian Chamber of Deputies, received on 3 November, on compliance with the principles of subsidiarity and proportionality by the proposed Cloud and AI Development Act regulation (COM(2026) 502 final, procedure 2026/0138(COD)). The opinion was adopted on 28 October by the XIV Committee (EU Policies) as Doc. XVIII-bis No 118. The committee found that the proposal complies with the principle of subsidiarity, subject to its observations, and only partly complies with the principle of proportionality. This is the national parliaments’ scrutiny provided for in Protocol No 2 to the Treaties, relevant for an act affecting cloud infrastructure and computing capacity for AI.
Council — doc. 15470/26 of 5 November
Dublin, 13 November: Heads of State or Government discuss AI
In his invitation letter of 3 November, European Council President António Costa convened an informal meeting of the Heads of State or Government in Dublin on 13 November, dedicated to artificial intelligence: the first discussion on this topic at the leader level. According to Eunews, the morning includes two sessions with external guests: with Arthur Mensch, CEO of Mistral AI, on competitiveness, sovereignty, critical dependencies, infrastructure, energy and investment; and with Geoffrey Hinton on the impact of AI on society, security and trust, including fundamental rights, child protection, democracy, market concentration, cybersecurity and the international governance of advanced systems. A working lunch of the 27 on the follow-up to the discussion will close the meeting.
European Council — invitation letter of 3 November
|
Irish Presidency — event page
|
Eunews — 5 November
---
EUROPEAN COMMISSION
2027 work programme and simplification report
On 4 November, the Commission adopted its 2027 work programme, together with the 2026 annual overview report on simplification, implementation and enforcement. For digital policy, there are two main announcements: a digital transformation package and a blueprint for the use of AI in five areas (health, transport, agri-food, manufacturing, and defence and space). In the security chapter, the Commission also commits to further assessing the impact of its data retention rules. On enforcement, the report records 874 infringement procedures launched, the highest number in ten years, 61 referrals to the Court of Justice, 25 requests for financial sanctions and 499 procedures closed after Member States complied. According to Commissioner Valdis Dombrovskis, the 12 omnibus packages and other measures put forward point to savings of up to EUR 22.3 billion a year, almost 60% of the target for the Commission’s term. The estimate assumes that the co-legislators adopt the proposals.
Commission — press release IP/26/2321 of 4 November
|
Commission — 2027 work programme and annexes
The Commission uses AI to draft and simplify laws
Presenting the simplification report, the Commission said it uses internal AI tools to draft new acts and simplify existing ones, helping staff identify overlaps, inconsistencies, duplications, and simplification opportunities. According to Politico, the Commission expects to sign a “pact against gold-plating” with Member States by the end of the year. The use of AI in lawmaking raises a non-trivial accountability issue: traceability of AI-generated contributions, effective human review, and handling bias in simplification suggestions will test the operation’s credibility.
Politico — 5 November
Age assurance: the independent study on fundamental rights and costs
On 4 November, the Commission published an independent study, carried out by an external contractor, on the impact of age assurance on fundamental rights, including data protection, privacy, non-discrimination and freedom of expression, and on the costs for businesses, including micro and small enterprises. The study does not assess the proposed EU KIDS Act or the compliance of any specific service. Its conclusions: an age assurance framework can be proportionate if accompanied by appropriate safeguards, including coherence with the GDPR, high privacy and security standards and the development of technical standards. The EU age verification blueprint, privacy-preserving, open by design and developed with public funding, could become one of the most affordable solutions for providers of all sizes; its rollout across all Member States should be accelerated, with fallback enrollment routes, including manual ones. Costs for online services are expected to remain modest for most operators and to decrease over time.
Commission — study of 4 November
Data Act: the Commission publishes model contract terms for data access and cloud
In Recommendation (EU) 2026/2394 of 29 October 2026, published in the Official Journal on 6 November, the Commission adopts non-binding model contractual terms on data access and use and non-binding standard contractual clauses for cloud computing contracts, implementing Article 41 of Regulation (EU) 2023/2854 (Data Act). The templates are designed to rebalance negotiations between the parties and reduce negotiation costs, particularly for SMEs and companies accessing data generated by connected products. The terms are optional: the parties may adopt them, adapt them to their needs or depart from them. The Commission notes, however, that they were drafted to be consistent with the rights and obligations under the Data Act and with each other, which should be borne in mind when modifying them.
EUR-Lex — Recommendation (EU) 2026/2394, OJ L of 6 November
Stronger education foundations in the age of AI
In Communication COM(2026) 770 final of 5 November 2026, “Stronger education foundations for all in the age of AI”, the Commission addresses three converging pressures on education systems. The first is the skills gap widened by AI: 40% of EU citizens still lack very basic digital skills. The second is the threat to democracy: 87% of citizens say online manipulation, including AI-generated content and deepfakes, threatens democratic processes. The third is minors’ early exposure: among 13-18 year-olds, 39% report encountering AI-generated content they find hard to recognize, 35% false or misleading information and 25% hate speech. PISA 2025 data confirm the picture: around one in three 15-year-olds underachieve in literacy and mathematics, and one in four in science. The Communication is accompanied by the staff working document SWD(2026) 775 final.
EUR-Lex — COM(2026) 770 final of 5 November
Review of the Digital Education Action Plan 2021-2027
SWD(2026) 775 final of 5 November, “Review of the Digital Education Action Plan (2021-2027) and Outlook on Digital Education”, accompanies the Communication and reviews the Plan’s implementation along its two strategic priorities: a high-performing digital education ecosystem and stronger digital skills. The actions assessed include the structured dialogue with Member States, the Council Recommendations on the key enabling factors for digital education and on blended learning, the European Digital Education Content Framework, schools’ digital transformation plans, the ethical guidelines on the use of AI and data in teaching, and the update of the European Digital Competence Framework to cover AI and data.
EUR-Lex — SWD(2026) 775 final of 5 November
Virkkunen reports on simplifying and enforcing the digital rulebook
On 4 November, Executive Vice-President Henna Virkkunen sent the Council and the Parliament her annual progress report on simplifying, implementing, and enforcing the digital rulebook. The achievements highlighted include the digital simplification package, with two omnibus proposals that, according to the Commission, chart a path for developing responsible AI, including when personal data are used, as well as the European Business Wallets. The report also covers the work with Member States on implementing and enforcing the rules.
Commission — news of 4 November
---
EUROPEAN PARLIAMENT
Child sexual abuse: LIBE approves the provisional agreement on the directive
On 4 November, the Committee on Civil Liberties, Justice and Home Affairs (LIBE) approved the text agreed in interinstitutional negotiations on the recast of the directive on combating the sexual abuse and sexual exploitation of children and child sexual abuse material (procedure 2024/0035(COD)). Coreper confirmed the agreement in a letter dated 23 October. The first-reading plenary vote is indicatively scheduled for 14 December 2026.
European Parliament — Legislative Observatory, procedure 2024/0035(COD)
EPRS: consolidating data protection rules for JHA bodies
On 4 November, the European Parliamentary Research Service published a briefing on the Commission proposal of 23 June 2026 consolidating data protection rules for EU bodies involved in law enforcement, judicial cooperation and border management. Building on a 2022 review that found the framework fragmented, the proposal amends Regulation (EU) 2018/1725 and recasts the Europol and Eurojust regulations. The changes, presented as largely technical, cover records of processing and operational logs, automated decision-making, data subject requests, breach notification and documentation, international transfers of operational personal data and the powers of the EDPS. This is the same package on which the EDPS issued three opinions in October.
European Parliament — EPRS briefing of 4 November
---
GARANTE PRIVACY ITALY
On 7 November 2026, the Garante issued a press release on the circulation on the dark web of personal data relating to Deputy Prime Minister Antonio Tajani and his family members. The message is addressed to media outlets, news organizations, websites and platforms: acquiring, disseminating or republishing by any means data originating from unlawful activities is itself a breach of data protection law, even when the data subject is a public figure. The authority announced that it will monitor the circulation of the data and, where necessary, urgently adopt measures to restrict or prohibit the processing and ask platforms to remove the content.
The warning is a reminder of the requirement that journalists publish only information essential to the story. The victim’s public profile alone does not justify republishing unlawfully obtained personal data. In its related-documents section, the page cites the decisions of 16 and 20 August 2025 in similar cases.
Garante — press release of 7 November
---
CNIL: FRENCH DATA PROTECTION AUTHORITY
Internal research in healthcare: the CNIL clarifies the boundaries
On 6 November 2026, the CNIL published a clarification on internal research carried out by health professionals. Research qualifies as “internal” only if it meets three cumulative conditions: it uses data collected in the individual care of patients, it is conducted by the staff providing that follow-up (the care team within the meaning of Article L. 1110-12 of the French Public Health Code), and it is for that staff’s exclusive use. In this configuration, no prior formalities with the authority are required, but the general GDPR obligations remain: records of processing, a DPIA where applicable, and information under Articles 12 to 14. If even one of the three criteria is not met, the controller must file a declaration of conformity with an existing reference framework or, failing that, obtain the CNIL’s prior authorization after the opinion of the competent committee. The declaration of conformity covers all the controller’s processing operations falling within the framework and need not be repeated for each study. On information, the CNIL recommends a layered approach, starting with individual information to the patient.
CNIL — news of 6 November
Session replay: the recommendation for providers and publishers
On 5 November 2026, the CNIL published its recommendation on session replay tools, adopted by deliberation No 2026-114 of 8 October following a public consultation. These trackers record and reconstruct users’ journeys on websites and apps (mouse movements, touch interactions, clicks, scrolling and, in some cases, data typed into forms). The text addresses both providers, who define the settings and features, and the publishers that deploy them. The authority stresses data minimization and the risk of collecting data not needed for the analysis, with guidance on characterizing the parties’ roles, informing data subjects, collecting consent under Article 82 of the French Data Protection Act, and technical and organizational measures, with particular attention to excluding sensitive fields from recording.
CNIL — recommendation published on 5 November
Privacy for Kids: the EDPB portal for children’s digital education
On 4 November, the EDPB launched Privacy for Kids, a portal bringing together educational resources from the data protection authorities of the EU, Iceland, Liechtenstein and Norway, including the CNIL. It targets parents, teachers and educators, with games, videos, quizzes and other materials that can be filtered by age group (3-10, 11-15 and 16+). Topics range from sharing online to social media, from smartphones and apps to privacy at home and at school, from cyberbullying to smart toys. The CNIL publicized the initiative alongside its own French-language resources for 8-10 and 11-15 year-olds.
EDPB — news of 4 November
|
CNIL — news of 4 November
CNIL plenary session of 5 November 2026
The session agenda includes a communication on fully automated individual decisions under Article 22 GDPR and the examination of a draft opinion on a decree amending the Code of Criminal Procedure concerning the FNAEG (genetic fingerprints) and FAED (fingerprints) databases. The agenda also lists the presentation of the CNIL/CIANum note “IA agentique et données personnelles”. For the guide “IA en contexte de soins”, it lists a hearing with representatives of the Haute Autorité de Santé and a presentation of the final version. Part II, under the simplified procedure, lists the examination of a draft opinion on a proposed processing operation titled “Analyse des dossiers de plainte déposés devant le Conseil supérieur de la magistrature”.
CNIL — agenda of 5 November
---
CASE LAW
Court of Justice of the EU, Case C-205/25, Lindenberg v Bayerisches Landesamt für Datenschutzaufsicht
5 November 2026 — Third Chamber — ECLI:EU:C:2026:933
In proceedings initiated by a complaint under Article 77 GDPR, the supervisory authority must be classified as a controller and is, in principle, required to grant the complainant access to the personal data concerning him or her and to the information listed in Article 15. Article 23(1)(c), (h) and (i) precludes a national provision, such as the one in the Bavarian Data Protection Act, that rules out entirely the complainant’s right of access to data held by the supervisory authority.
InfoCuria
Court of Justice of the EU, Case C-468/24, SR v Netz Niederösterreich GmbH
5 November 2026 — Fourth Chamber — ECLI:EU:C:2026:936
Smart electricity meters. The grid operator is not required to accommodate a final customer who refuses a smart meter, nor to supply a conventional meter instead (Directive (EU) 2019/944). A national definition of a smart meter with no specific data security requirement is permissible if other national rules implement the data security and data protection requirements of the directives. Article 5(3) of the ePrivacy Directive does not apply to an electricity network through which personal data may be transmitted. The question on the interpretation of the GDPR was declared inadmissible. Text available in German and French.
InfoCuria
Court of Justice of the EU, Case C-250/25, Like Company v Google Ireland — Opinion of Advocate General Szpunar
3 November 2026 — ECLI:EU:C:2026:919
The first generative AI case before the Court (Gemini chatbot). The Advocate General proposes declaring the questions on model training inadmissible as hypothetical, while observing that training is, as a rule, covered by the text and data mining exception. Real-time retrieval and reuse of press articles to generate answers (RAG) may, by contrast, infringe press publishers’ rights in certain circumstances: what matters is unlawful access, for instance by circumventing a paywall, and the recognizable reproduction of the form of expression to the point of sparing the user from reading the original. In the Advocate General’s view, the provider may be liable if such use is inherent to its business model or if it has not put in place credible and effective measures to counter it.
Court of Justice — press release No 155/26
Court of Justice of the EU, Case C-654/25 — Opinion of Advocate General Szpunar
5 November 2026 — ECLI:EU:C:2026:948
A reference from the Bundesgerichtshof on the dynamic embedding of Google Fonts in a website, which sends the visitor’s IP address to Google. The Advocate General proposes that a dynamic IP address be classified as personal data for the third-party recipient, where it has means reasonably likely to be used to identify the person, and for the website operator, indirectly, where the operator knew or should have known this. Under Article 82 GDPR, no non-material damage can be recognized for a person who knowingly provoked the infringement, in an automated and repeated way, in order to obtain compensation. Text available in French and German.
InfoCuria
We have no ECtHR judgments under Article 8 of the Convention concerning data protection or surveillance to report this week.
---
TikTok’s “For You” rabbit hole under Commission scrutiny
On 6 November, the European Commission sent TikTok its preliminary findings: the recommender system of the “For You” feed would breach the Digital Services Act because of its “rabbit hole” effect. The more a user engages with a certain type of content (sadness, isolation, body image), the more the algorithm recommends similar material, with potentially harmful effects on the mental health of vulnerable adults and minors. The objections concern the two pillars of the obligations of very large online platforms: systemic risk assessment and mitigation measures. These are preliminary findings: TikTok can reply before any non-compliance decision, which could lead to fines of up to 6% of worldwide turnover.
Commission — press release IP/26/2361 of 6 November
|
Commission — news of 6 November
Call for evidence on advertising repositories under Article 39 DSA
On 4 November, the Commission opened a call for evidence, closing on 2 December 2026, on technical guidelines for the advertising repositories required by Article 39 DSA. The guidelines will address three aspects: searchability of the repositories, information on targeting and reach criteria, and accessibility of ad content. The stated aim is to make the repositories of VLOPs and VLOSEs genuinely usable for public scrutiny, public health, consumer protection and electoral integrity, as well as for the Commission’s own supervision. A targeted public consultation is planned for 2027, aimed at researchers, civil society, platforms and the advertising industry, with the guidelines due in the second quarter of 2028. Those handling DSA compliance on the advertising side have a narrow window to influence the technical design.
Commission — call for evidence of 4 November
The NYT and other papers sue Brave over scraping and paywall circumvention
The New York Times has sued Brave Software and, in a separate action, the New York Daily News, the Chicago Tribune and the Orlando Sentinel have done the same. The publishers accuse Brave of unlawfully extracting content from their websites and reselling real-time access to the “Brave Index” to AI companies, including Anthropic, Perplexity, Mistral, xAI and Cohere, which allegedly use it to train models and answer user queries. The second strand concerns the browser’s “Paywall blocker” feature, which allegedly lets users bypass paywalls without authorization or compensation. Steven Lieberman, counsel for the three regional papers, describes the operation as “laundering” content for the AI industry. Building the claim on two levels, direct infringement through scraping and downstream distribution and the neutralization of commercial protection measures, shifts the litigation from model developers to the intermediaries supplying the data. The issue connects with the Advocate General’s Opinion in Case C-250/25, reported under Case Law.
The Verge — 2 November
---
EUROPEAN STANDARDIZATION
CEN and CENELEC introduce the European Agile Specification
On 6 November, CEN and CENELEC announced the approval of the European Agile Specification (EAS), a new European standardization deliverable designed to bring shared technical solutions to market faster. The EAS will become a permanent part of the two organizations’ toolbox from 1 January 2027: rather than waiting until every aspect of a specification is finalized, it allows mature content to be published and then developed through successive iterations. According to the pilots, an EAS can be developed in less than a year. The decision implements the Cyprus Commitment, through which the 43 members of CEN and CENELEC undertook to make standardization faster, more digital and more open, while preserving consensus and national delegation. For fast-moving fields such as AI, in my view, it remains to be seen how an iterative instrument will fit with the harmonized standards required by EU legislation.
CEN-CENELEC — news of 6 November
---
INTERNATIONAL DEVELOPMENTS
State Aging Plans and AgeTech in the United States
On 5 November, the Future of Privacy Forum published an analysis by Jordan Wrigley on the role of State Aging Plans, the multi-year strategic plans that US states and territories must adopt to access federal funding under the Older Americans Act, as a junction between the development of technologies for older adults and the protection of personal data. The plans steer the allocation of OAA funds, American Rescue Plan resources and Medicaid home and community-based services waivers through the network of Area Agencies on Aging.
The analysis distinguishes three approaches: states that integrate AgeTech as a pillar of public service (assistive technology lending, tablets, robotic pets, digital literacy), states that keep traditional non-digital models, and states that contract private vendors directly. One example is the Washington State Department of Social and Health Services, which works with Intuition Robotics to provide the AI-based ElliQ robot to eligible Medicaid recipients.
The relevant point for privacy and consumer protection professionals is that public procurement moves the governance of vulnerable people’s data into heterogeneous state procedures, with protection standards that differ across jurisdictions.
FPF — 5 November
UN experts on the independence of the Irish DPC
In a statement issued in Geneva on 4 November, two UN Special Rapporteurs, Ana Brian Nougrères (right to privacy) and Margaret Satterthwaite (independence of judges and lawyers), raised concerns about the independence of the Irish Data Protection Commission and about financial barriers to access to justice, pointing to the risk of undue influence by large platforms and saying they are in contact with the Irish government. The issue carries weight because Apple, Google, Meta and Microsoft have their European headquarters in Ireland, making the DPC the lead GDPR authority for much of the US technology industry.
The Rapporteurs call for regulator selection processes based on objective, specialist criteria, impartial selection panels and the prior identification of conflicts of interest, with disclosure and recusal obligations where necessary. According to Politico, the implicit reference is to the September 2025 appointment of former Meta lobbyist Niamh Sweeney as commissioner. DPC spokesperson Evan O’Leary rejected the criticism, pointing to fines of more than EUR 4.4 billion and to the ethics rules binding the commissioners. The underlying issue remains political: the one-stop-shop mechanism concentrates contested authority in decisions that take effect across the whole Union.
Politico — November 2026
---
ARTIFICIAL INTELLIGENCE
ChatGPT Lockdown Mode bypassed: silent exfiltration of chats
On 5 November, Check Point disclosed a now-patched vulnerability that allowed attackers to bypass Lockdown Mode, which OpenAI introduced in June 2026 to reduce prompt-injection risk. The flaw lay in the HTML preview mechanism for code blocks: nested sandboxed frames allowed an unintercepted popup function to run, bypassing restrictions on outgoing network requests. Check Point demonstrated that a malicious GPT could exfiltrate chat messages without a click in the ChatGPT Classic desktop app. The researchers also described how the flaw could be exploited for phishing or malware delivery. OpenAI fixed the issue on 18 September 2026. The case reminds us that controls on the model are not enough when the attack surface is the application layer around it.
The Hacker News — 5 November
CERT-In reports an ongoing campaign by the threat actor Crafty Falcon against banks, insurers, cryptocurrency exchanges and fintechs in India, using publicly available AI-based offensive tools such as ARTEX, PentAGI, Cairn and CyberStrikeAI. Attack chains run autonomously, at machine speed, with scripts generated and adapted to the target’s responses and infrastructure rotating IP addresses within hours or days. Initial access relies on stolen static API tokens, repository compromises through forged commits, and plaintext credentials in configuration files and services without MFA. Once inside, the attackers call fund-transfer APIs directly from the victim’s authorized infrastructure, bypassing application-level controls on KYC, transaction limits and beneficiary verification. The alert follows one issued in October 2026 on NBFCs and payment service providers.
The Hacker News — 5 November
Prompt adware: advertising injected into an OpenAI Codex jailbreak
Check Point Research analyzed “Haiou” (Seagull), a jailbreak bundle for OpenAI Codex distributed on GitHub as “haiou-codex” (403 stars and 77 forks since 10 July 2026), which modifies the agent’s global configuration files to inject persistent advertising for a third-party LLM gateway (zz.aiapi2025[.]top). Trigger words such as “hi”, “hello” or “Seagull” bring up the ad. The jailbreak combines role-play, reframing requests as CTF scenarios and installing skills for pentesting, cracking and reverse engineering. The bundle also ships with Codex-X, a Chinese graphical interface for configuring Codex, and researchers suspect the gateway recycles stolen LLM tokens. The business model introduces a new category: adware delivered through natural-language instructions rather than an executable, and therefore invisible to traditional payload checks.
The Hacker News — 4 November
VOLTA-AUTODOXER: the infostealer that uses Claude, Gemini and Kimi to price its victims
Cisco Talos has identified VOLTA-AUTODOXER, a Windows infostealer that steals the usual (Chromium browser passwords and cookies, Outlook profiles, Slack workspaces, payment cards, wallet seeds, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud keys, Wi-Fi profiles, screenshots) but adds an AI-based profiling layer. The malware builds a structured dossier with geolocation, home address (confirmed by reading e-commerce and delivery receipts in the victim’s Gmail), employer, presumed year of birth and the last four digits of payment cards. The dossier is sent through OpenRouter to Claude Opus 4.6, falling back to Kimi K2 and Gemini 2.5 Flash if the model refuses: refusals are detected by looking for phrases such as “I can’t”. A 240-word system prompt asks the model to act as a “forensic intelligence analyst” and produce a dossier with sections titled “WHAT THEY’RE WORTH” and “THEIR OPSEC”. Exfiltration runs over Telegram in three stages. What matters is not the stealer, but the use of commercial LLMs as a victim-assessment service, switching from one model to another to bypass individual providers’ refusals.
The Hacker News — 2 November
Australia: Centre for Public Integrity fears a new robodebt from automated decisions
The “Computer Says No” report by the Centre for Public Integrity, launched on 5 November by CEO Catherine Williams, accuses the Albanese government of sleepwalking towards a new robodebt because of the lack of transparency and rules on the use of algorithms and AI in administrative decisions. It makes eight recommendations, including a mandatory audit, a public register of all automated decision-making systems in use and an independent body with monitoring and investigative powers. The case that triggered the report is the Integrated Assessment Tool used to assess older people’s entitlement to home care: the tool allegedly underestimates needs, and assessors are reportedly prevented from correcting its outputs. The Department of Attorney-General Michelle Rowland is working on an automated decision-making framework (consultation closed in January 2025) meant to implement the recommendations of the Robodebt Royal Commission, with no set timeline for a draft. The crux is using the tools operationally while the rules remain under construction.
The Guardian — 5 November
Facial recognition in Western Australia: 1.38 million faces scanned, 122 arrests
Western Australia Police have concluded a five-month trial of real-time facial recognition, with cameras mounted on vans at 43 locations, including the Perth Royal Show. Across 100 deployments, 1,380,882 faces were scanned, leading to 122 arrests, contact with 146 registered sex offenders, 10 welfare checks and 9 false alerts. Commissioner Col Blanch presented an error rate of 0.00065% of all scans. Bill Rowlings of Civil Liberties Australia disputes the metric: the rate should be measured on arrests compared with actual convictions, not on everyone scanned. The objection touches on a recurring methodological point in European debates on biometrics, and Article 5 of the AI Act: the choice of denominator changes the order of magnitude of the risk and shapes the proportionality assessment.
The Guardian — 2 November
---
CYBERSECURITY
Refrigeration systems under attack: the EU joint statement and the FBI-CISA advisory
On 4 November, ENISA, the Commission (DG CNECT), Europol’s European Cybercrime Centre (EC3), the CSIRTs Network and EU-CyCLONe issued a joint statement on the incidents that, since 11 October, have hit commercial refrigeration systems and similar devices in several Member States. After a new wave on 1 and 2 November, organizations are urged to reduce the exposure of the affected systems immediately. The statement recalls that since 11 September 2026, manufacturers of products with digital elements must report severe incidents and actively exploited vulnerabilities through the Cyber Resilience Act’s single reporting platform, and that NIS2 entities must manage cybersecurity risks and notify significant incidents.
In the United States, a joint advisory by the FBI, CISA, HHS, USDA and DOT describes attacks on unpatched or internet-exposed temperature controllers and refrigeration systems. Attackers bypassed authentication, injected code or took advantage of weak configurations to change setpoints, trigger defrost cycles or impair temperature monitoring. The advisory refers to the European cases: between 250 and 350 French supermarkets hit in October, more than 115 stores in six countries in November with temperatures driven up to 50 °C and, on 3 November, confirmation by the Établissement français du sang that settings had been altered at three blood storage facilities through an equipment supplier. The recommendations are the classic OT ones: remove controllers from the public internet, change default passwords, patch, and separate OT from IT networks.
ENISA — joint statement of 4 November
|
BleepingComputer — 5 November
ENISA: technical guidance on secure update mechanisms
On 4 November, ENISA published a Technical Advisory on secure update mechanisms, aimed at manufacturers of products with digital elements and in particular the teams that design and operate updates. The document is intended mainly for micro, small and medium-sized enterprises: it describes threats across the update life cycle and a set of controls to mitigate them. It is practical support ahead of the Cyber Resilience Act requirements, which will apply from 11 December 2027 and include remediating vulnerabilities without delay throughout the support period.
ENISA — Technical Advisory of 4 November
ECCC: seven projects for the cybersecurity of hospitals and healthcare providers
On 6 November, the European Cybersecurity Competence Centre (ECCC) launched seven projects funded by the Digital Europe Programme for about EUR 30 million, in line with the European action plan on the cybersecurity of hospitals and healthcare providers. The projects aim to deliver more than 200 tools, services and methodologies, involve more than 500 healthcare institutions with over 200 pilot installations, train at least 10,000 staff and support compliance with NIS2, the GDPR and the European Health Data Space. Pilots are already scheduled in 14 Member States.
ECCC — news of 6 November
Iris Ohyama: up to 7.7 million people potentially affected by a third-party cloud incident
Japanese manufacturer Iris Ohyama confirmed unauthorised access to an external cloud service used by Iris Plaza. The service held records for about 7.7 million people, including 5.8 million current and former members and 1.9 million non-members. The company was still investigating how many records had actually been accessed or taken. It suspended the site on 28 October after finding a suspicious file and notified Japan’s Personal Information Protection Commission and the police on 29 October. The data potentially involved include names, addresses, phone numbers and email addresses. Member records may also include gender, dates of birth and, for current members, encrypted passwords. The company confirmed that credit card data were unaffected and said it had received no reports of misuse or financial loss linked to the incident.
BleepingComputer — 8 November
QNAP: the Plex2Shell chain exploited against exposed NAS devices
In an advisory of 6 November, QNAP reports suspected compromises of NAS devices running Plex Media Server versions vulnerable to the Plex2Shell chain (CVE-2026-96652, 96654, 96655, 96656), discovered by Zach Main and reported to Plex on 27 July. Exploitation began after technical details were published on 16 October; on 20 October, CISA added the four CVEs to its KEV catalogue with a three-day deadline for federal agencies. The fix is in Plex version 1.43.3.10861; QNAP operating systems are not vulnerable. The vendor is explicit: treat NAS devices as potentially compromised if they were reachable from the internet while running a vulnerable version of Plex, because the update does not remove payloads already installed. The advice: check the indicators of compromise, reinstall the system, reset Plex and NAS credentials, close port forwarding on port 32400 and prefer remote access via VPN.
BleepingComputer — 6 November
Crédit Mutuel Arkéa: USD 133 million stolen through a service provider
The French cooperative bank Crédit Mutuel Arkéa (5.8 million members, EUR 211.5 billion in assets) suffered an attack that generated fraudulent transfers estimated at EUR 118 million (USD 133 million): the bank confirms the order of magnitude, not the exact figure. The attacker entered the network through a service provider’s platform (Fortuneo has been ruled out) and operated from the bank’s own account towards other institutions, then converted the funds into USDC and USDT. The public alarm was raised on 2 November by investigator ZachXBT; the bank disclosed the incident on 4 November. According to ZachXBT, the delay allowed about USD 30 million to be laundered before European authorities could freeze it. Customer assets are not reported to be affected, internal human error has been ruled out, and judicial authorities are leading the investigation. Once again, the effective perimeter coincides with that of the providers with privileged access.
BleepingComputer — 7 November
WorkOS: the researcher says an AI agent escaped its sandbox
WorkOS attributes the breach of its internal sales application to an agent of the ARTEX offensive framework, operated by a researcher who says the tool escaped its test environment and unintentionally interacted with the company’s systems. The researcher came forward after the first notice; WorkOS and an independent forensic consultant matched timestamps, request details and technical artefacts against the logs, confirming the researcher’s responsibility. How the sandbox escape happened has not been explained. The exfiltrated data were reportedly deleted and not shared. Technically, the agent exploited four cumulative weaknesses: a legacy platform lacking the protections of newer systems, internet-exposed API endpoints with permissive access controls, a predictable credential intended for local development and left in production, and no request rate limiting, which allowed records to be extracted quickly. First access on 25 October, discovery on 30 October. Setting aside the sandbox story, this remains a breach caused by development credentials left in production: AI amplifies the speed, it does not create the vulnerability.
BleepingComputer — 6 November
Japan: a JPY 10 billion emergency cybersecurity package
After a series of breaches affecting companies and public bodies, the Japanese government announced a JPY 10 billion (USD 63 million) emergency package. Prime Minister Sanae Takaichi called for a step change in capabilities against AI-generated attacks. The plan uses the “active cyber defense” legalized in October, strengthening preemptive access to servers used by attackers; a joint team from the Ministry of Defense, the National Police Agency and the National Cybersecurity Office for threats to national security; a METI “top gun team”, operational by the end of November, that will use AI to find vulnerabilities and counter hostile actors in cooperation with technology companies and AI developers from allied countries; and a public-private “cyber rescue team” for first response to victims.
SecurityWeek — 6 November
Frontline Education: more than 1.3 million people affected
Frontline Education, a software provider for US school districts, is notifying people affected by a data breach. At least 1,298,917 people were affected, according to filings with the attorneys general of Massachusetts, New Hampshire, South Carolina, Texas and Vermont; the real figure is likely higher, given notifications in California, Indiana and Iowa as well. The attack, which exploited a vulnerability in third-party software, took place on 14 August and was contained the following day. The information taken from Frontline Central and School Health Management varies by individual and may include name, Social Security number, date of birth, driver’s license number, passport and other identity documents, health insurance data (including Medicaid identifiers) and student identifiers. Two years of credit monitoring are offered to adults and dark web monitoring to minors.
SecurityWeek — 3 November
Barracuda: two chainable flaws leading to root
Barracuda has disclosed two vulnerabilities that, chained together, lead to an authorization bypass and root access: CVE-2026-116606 (CVSS 9.3), a pre-authentication bypass limited to Email Security Gateway, and CVE-2026-111560 (CVSS 8.6), a command injection affecting a wide range of products: Email Security Gateway, Load Balancer, Application Delivery Controller, Message Archiver, WAF Control Center, Web Application Firewall and Web Security Gateway. Exploitation requires access to the administration interface, so installations that do not expose it to the internet are at lower risk. The advice: update to the latest firmware, with several successive updates if the version is very old, and restrict access to the administration interfaces to trusted internal networks.
The Hacker News — 3 November
---
TECH & INNOVATION
SonicWall: SAML bypass on SSL VPN already exploited
On 6 November SonicWall released a patch for CVE-2026-19868, a flaw in the SonicOS SSL VPN service that lets a user with valid credentials for an alternate authentication method bypass SAML controls, including SAML-enforced MFA. CVSS 7.1, reported by Ryan Harris of The Electronic Office. The vendor confirms it investigated possible active exploitation; CISA added it to the KEV catalog the same day, with a 20 November deadline for federal agencies under BOD 26-04.
Exposure arises only when SAML login is enabled. Affected are Gen7 TZ, NSa, NSsp and NSv firewalls running SonicOS 7.3.3-7015 or earlier and Gen8 TZ, NSa and NSv running 8.2.2-8015 or earlier. Fixes are in versions 7.3.4-7013 and 8.2.3-8015, and no workarounds exist.
The advisory covers six further vulnerabilities not flagged as exploited: two in the management interface (CVSS 8.8, a permissive cross-domain policy and a session validation flaw) usable without authentication, three pre-authentication buffer overflows in the SSL VPN service and the TACACS+ client exploitable for application-level DoS, and weak password hashing in virtual appliances. The KEV catalog now lists twenty SonicWall vulnerabilities, but not yet CVE-2026-102255, under attack since last month.
SecurityWeek — 7 November
OpenSSH on musl: oversized passwords unlock password-disabled accounts
On 5 November, Ozgun Kultekin disclosed the “*sshole” vulnerability, which allows an attacker to log in to affected Linux accounts whose password hash has been replaced by an asterisk in /etc/shadow, using a password longer than 256 bytes. The flaw has existed since 2016 and stems from the interaction between the OpenSSH patch for CVE-2016-6210, which forces password hashing even for accounts without a usable hash to equalize response times, and musl’s crypt() function, which returns * on error (including an overlong password). The returned * matches the value in /etc/shadow, so authentication succeeds.
Vulnerable are OpenSSH installations on musl-based distributions: Alpine Linux, OpenWrt, Buildroot, Yocto. Required conditions: PAM disabled (the default in Alpine’s openssh-server package), SSH password authentication enabled, at least one account with * as its hash and another with a valid hash. glibc systems (Debian, Ubuntu, Fedora) and PAM-enabled setups are not affected. Alpine’s official cloud images for AWS and Azure expose the alpine account with passwordless doas privileges, a scenario that leads to root. The fix is in OpenSSH 10.6p2; on Alpine, apk upgrade openssh.
Kultekin found the flaw through about 17,600 authentication tests orchestrated with AI agents across configurations, malformed hashes and oversized passwords.
BleepingComputer — 5 November
On 4 November, the Commission launched the European Chip Design Platform, a Chips Act initiative that will mobilise EUR 525 million in investment to foster a new generation of European chip design companies. The platform gives start-ups and SMEs access to design tools, design support services, intellectual property blocks and manufacturing capacity: the resources that are most expensive and hardest to obtain at an early stage. European chip design companies account for less than 1% of global revenues in the segment. The Commission argues that new manufacturing capacity in Europe depends on demand from European designers. For those following technological sovereignty, it adds a piece to the debate on critical dependencies that Heads of State or Government will take up in Dublin.
Commission — press release IP/26/2345 of 4 November
|
Commission — news of 4 November
---
SCIENTIFIC RESEARCH
Selection of the most relevant papers of the week from arXiv on AI, Machine Learning and Privacy
Memorization and Unlearning in LLMs
An Information Bottleneck for Post-hoc Memorization Mitigation in LLMs
Submitted on 5 November 2026; announced on 6 November 2026.
The authors propose a post-hoc approach based on the information bottleneck to reduce memorization of copyright-protected training data, without modifying or retraining the underlying model. It is interesting because it opens the way to interventions on models already in production, where retraining is not economically or technically feasible.
arXiv
Forget What Comes Next: LLM Unlearning via Statistical Independence
Submitted on 2 November 2026; announced on 4 November 2026.
The paper shows that poisoned examples remain statistically detectable even after current unlearning techniques are applied, casting doubt on whether these methods really remove what the model has memorized. The authors propose an unlearning objective based on statistical independence. Relevant to the debate on the effectiveness of Article 17 GDPR applied to language models: declared “erasure” may not be erasure at all.
arXiv
Fewer Heads, Fewer Memories: Pruning Redundant Attention Heads Suppresses Verbatim Memorisation
Submitted on 29 October 2026; announced on 2 November 2026.
Starting from the hypothesis that attention heads work as associative memories, the authors show that pruning redundant heads reduces verbatim reproduction of training sequences. An architectural approach to mitigation, alternative both to data filtering and to interventions at generation time.
arXiv
Differential Privacy: Algorithms and Accounting
Privacy for Free in Efficient Agnostic Halfspace Learning
Submitted on 3 November 2026; announced on 4 November 2026.
The first computationally efficient, differentially private algorithm for agnostically learning halfspaces, analyzed in three settings (improper learning under Gaussian and log-concave distributions, proper Gaussian learning, semi-agnostic Gaussian). The “privacy for free” in the title means the DP guarantee does not appreciably worsen the result compared with the non-private case.
arXiv
Fast Privacy Loss Accounting for Random Allocation via Multiscale FFT
Submitted on 30 October 2026; announced on 2 November 2026.
The authors speed up privacy loss computation (PLD accountant) for the random allocation sampling scheme, in which each data point is used in k steps chosen uniformly out of t, improving on the previous algorithm that was quadratic in the inverse of the accuracy parameter. A useful result for anyone calibrating the ε budget in real private optimization or federated aggregation systems.
arXiv
LLM Agents, Contextual Privacy and Reliability
PrivacyFlow: Privacy in AI Agents through Information Flow Analysis and Consequence-Aware Risk Management
Submitted on 29 October 2026; announced on 2 November 2026.
Autonomous LLM agents running multi-step workflows on personal data tend to drift from users’ privacy expectations and pass information out of context. PrivacyFlow operationalizes Contextual Integrity through information flow analysis and consequence-aware risk management, in realistic sequences and dynamic environments. Useful material for DPIAs on agentic systems, where the scope of processing is by definition shifting.
arXiv
When AI Investigators Cannot Trust Their Tools: Conflict-Aware Evidence Verification for Cyber Threat Investigation
Submitted on 30 October 2026; announced on 2 November 2026.
LLM agents that analyze security alerts combine the outputs of several detection tools, implicitly treating them as reliable. The paper studies what happens when the tools produce conflicting or wrong evidence and proposes a verification mechanism that accounts for conflicts. An underestimated problem in automated SOCs, where the digital chain of evidence can inherit unflagged errors.
arXiv
Federated Learning in Critical Settings
Communication-Efficient Private Federated Reinforcement Learning with Post-Quantum Authentication for Vehicular Networks
Submitted on 29 October 2026; announced on 2 November 2026.
A federated reinforcement learning framework for cooperative vehicle routing that tackles three simultaneous constraints: privacy of updates, communication efficiency over intermittent vehicle-to-infrastructure links, and post-quantum authentication, evaluated in simulation. The combination is specific to the automotive setting, where security, latency and data protection force non-trivial trade-offs.
arXiv
---
AI ACT IN A NUTSHELL: Part 46
Article 50 - Transparency obligations for providers and deployers of certain AI systems
In Part 45 we examined Article 49, on registration. We now turn to Article 50, which introduces a cross-cutting pillar of the AI Act: transparency obligations for providers and deployers of certain AI systems, regardless of whether they are classified as high-risk. The obligations have applied since 2 August 2026, and Regulation (EU) 2026/1744 amended Article 50(7) and added a transitional rule for certain providers under Article 111(4).
Article 50 identifies four distinct scenarios, with responsibilities allocated differently along the AI value chain.
The first concerns AI systems intended to interact directly with natural persons, such as chatbots and conversational assistants. Providers must design and develop them so that people are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context. The obligation does not apply to systems authorized by law to detect, prevent, investigate or prosecute criminal offences, subject to appropriate safeguards for the rights of third parties, unless those systems are available to the public to report a criminal offence.
The second involves providers of AI systems, including general-purpose AI systems, generating synthetic content (audio, images, video or text). Outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities of the content, the costs of implementation and the generally acknowledged state of the art. Excluded are assistive functions for standard editing and systems that do not substantially alter the input data provided by the deployer or its semantics, as well as uses authorized for law enforcement purposes.
The third is addressed to deployers of emotion recognition or biometric categorization systems, who must inform the people exposed about how the system operates and process personal data in accordance with the GDPR, Regulation (EU) 2018/1725 and Directive (EU) 2016/680. Here too there is an exception for systems authorized by law to detect, prevent or investigate criminal offences, with appropriate safeguards. The explicit reference to data protection law underlines the cumulative nature of the obligations.
The fourth concerns deployers generating or manipulating deep fakes and those publishing generated or manipulated text to inform the public on matters of public interest: in both cases the artificial nature of the content must be disclosed. For evidently artistic, creative, satirical or fictional works, the obligation is limited to disclosing the existence of generated content in an appropriate manner that does not hamper the enjoyment of the work. For text, the obligation does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it.
Paragraph 5 requires the information to be provided in a clear and distinguishable manner, no later than the first interaction or exposure, in line with accessibility requirements. Paragraph 6 states that paragraphs 1 to 4 do not affect the requirements of Chapter III or other transparency obligations under Union or national law.
The amendments made by Regulation (EU) 2026/1744
Paragraph 7 has been replaced. In the original text, the AI Office encouraged codes of practice on detection and labelling, and the Commission could approve them by implementing act. Now it is the Commission that encourages and facilitates Union-level codes on the detection, marking and labelling of artificial content and assesses, taking utmost account of the Board’s opinion, whether adherence to the codes is adequate for the obligations in paragraphs 2 and 4, under the procedure in Article 56(6). The power to approve codes by implementing act has been removed; the Commission may still adopt common rules by implementing act if a code is deemed inadequate. Recital 41 of Regulation 2026/1744 clarifies that the codes do not confer a presumption of conformity.
A transitional period has also been introduced (new Article 111(4)): providers of systems generating synthetic content placed on the market before 2 August 2026 must comply with the marking obligation in paragraph 2 by 2 December 2026. The other paragraphs of Article 50 benefit from no extension.
On 20 July 2026 the Commission published its guidelines on the transparency obligations for providers and deployers of certain AI systems, to be read alongside the provision.
Practical implications
For DPOs and compliance officers, Article 50 requires mapping AI use cases beyond risk classification: systems that are not high-risk may still fall under the transparency obligations. Organizations need to establish whether they act as provider, deployer or both, because the obligations are allocated differently.
Providers subject to Article 50(2) must mark synthetic outputs in a machine-readable format and make them detectable as artificially generated or manipulated. Possible techniques include watermarking, metadata and fingerprinting. Documenting the choice can help show how the provider took account of the state of the art. Providers whose systems were placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2). Adhering to a code assessed as adequate helps, but does not replace verifying compliance. Deployers of deep fakes and informative text must set up disclosure processes, assessing whether the exceptions for creative works or for content subject to editorial review with an assumption of responsibility apply.
For emotion recognition and biometric categorization systems, the information required by Article 50 is cumulative with GDPR obligations: legal basis, transparency, DPIA and proportionality assessment remain central. Contractually, clauses with suppliers and integrators should be updated to allocate correctly the obligations on marking and on informing end users, avoiding grey areas of responsibility.
Legislative sources:
AI Act
|
Regulation (EU) 2026/1744
|
Commission — guidelines of 20 July 2026
Next installment
In Part 47 we turn to Article 51 - Classification of general-purpose AI models as general-purpose AI models with systemic risk.
---
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 implementing them 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 fully functional original Python tools reproduced in full (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 English and Italian (Italian title: Agentic AI — Autonomia tecnica, responsabilità umana e governance giuridica).
Find out more →
The book page
---
FROM THE NICFAB BLOG
What Can a Sender Learn When You Open an Email?
6 November 2026
How a tracking pixel works, what it reveals to the sender, what the law says and how to block it in Gmail, Apple Mail, Outlook, Thunderbird and Proton Mail.
Read the full article
Algorithms and training data as trade secrets: the amendment to the Industrial Property Code in Legislative Decree 179/2026
5 November 2026
Italian Legislative Decree 179/2026: data, algorithms and mathematical methods for training AI systems as confidential business information, Article 98 of the Industrial Property Code, Legislative Decree 160/2026 and the AI Act.
Read the full article
NIS2, GDPR and DORA: first the applicable regime, then the coordination
4 November 2026
Relationship between NIS2, GDPR and DORA in Italy: Article 4 of the Directive, exclusions for financial entities, dual notification, Article 14 of Legislative Decree 138/2024, the perimeter and Law 90/2024.
Read the full article
AI and employment in Legislative Decree 179/2026: automated decisions, reasons, safety and nullity
2 November 2026
AI and employment in Italian Legislative Decree 179/2026: mandatory human decision, performance evaluations, consent, reasons, nullity, exclusion of screening, risk assessment, employment services, GDPR and AI Act.
Read the full article
---
Featured events and meetings
Web Summit 2026, with the Commission’s initiatives in Lisbon (scheduled: 10-12 November 2026)
European Commission |
Info
IMCO Committee: presentation of the EU KIDS Act and exchange of views on advanced AI systems (scheduled: 10 November 2026)
European Parliament |
Info
Presentation of ENISA’s report on the State of Cybersecurity in the Union to the ITRE Committee (scheduled: 10 November 2026)
European Parliament |
Info
Council Working Party on Telecommunications and Information Society: first Presidency compromise text on the Cloud and AI Development Act (Title IV, Autonomy, and Annexes II and III) (scheduled: 10 November 2026)
Council of the European Union |
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
European Parliament plenary: debate on the cyber resilience of critical infrastructure and AFCO report on the institutional aspects of artificial intelligence (scheduled: 11-12 November 2026)
European Parliament |
Info
Hearings in Cases C-744/25 P (Commission v Meta) and C-745/25 P (Commission v TikTok) on the DSA supervisory fee; Advocate General’s Opinion in Case C-826/24 P (Google AdSense) (scheduled: 12 November 2026)
Court of Justice of the European Union |
Info
HMA/EMA multistakeholder AI and Data Forum: AI and data in medicines regulation, Amsterdam (scheduled: 12-13 November 2026)
EMA |
Info
Informal meeting of the Heads of State or Government on artificial intelligence, Dublin (scheduled: 13 November 2026)
European Council |
Info
Working Party on Data Protection: revision of Regulation (EU) 2018/1725 (scheduled: 13 November 2026)
Council of the European Union |
Info
Public consultation on EDPB Guidelines 04/2026 on the application of the power to impose administrative fines in relation to other corrective powers (deadline: 13 November 2026, 23:59 CET)
EDPB |
Info
air 2026 — Convaincre ou manipuler ? L’éthique du numérique à l’épreuve des urnes (scheduled: 16 November 2026)
CNIL |
Info
Grand Chamber judgment in Case C-698/23 P, EDPS v Parliament and Council (Europol Regulation) (scheduled: 17 November 2026)
Court of Justice of the European Union |
Info
Apply AI Summit (scheduled: 17 November 2026)
European Commission |
Info
Advocate General’s Opinion in Case C-643/25, Verbraucherzentrale Bundesverband (“free” digital service) (scheduled: 19 November 2026)
Court of Justice of the European Union |
Info
EYCS Council: conclusions on skills and AI and on young people and the information environment (scheduled: 19 November 2026)
Council of the European Union |
Info
Telecommunications Council (scheduled: 1 December 2026)
Council of the European Union |
Info
Journée RGPD 2026 in Poitiers (scheduled: 1 December 2026)
CNIL |
Info
Call for evidence on advertising repositories under Article 39 DSA (deadline: 2 December 2026)
European Commission |
Info
Hearing in Case C-730/25, Vinted (scheduled: 2 December 2026)
Court of Justice of the European Union |
Info
Feedback on the proposal for an EU KIDS Act regulation (deadline: 3 December 2026)
European Commission |
Info
Safer Internet Forum 2026 (scheduled: 3 December 2026)
European Commission |
Info
Product Cybersecurity: Making CRA a Reality (scheduled: 11 December 2026)
ENISA |
Info
Plenary vote on the child sexual abuse directive (indicative date) (scheduled: 14 December 2026)
European Parliament |
Info
---
Conclusion
Three reports from the past week show how AI tools are being used in cyberattacks. Cisco Talos documents VOLTA-AUTODOXER, an infostealer that routes victims’ data to Claude Opus 4.6, falling back to Kimi K2 and Gemini 2.5 Flash when the model refuses, to produce dossiers complete with a “WHAT THEY’RE WORTH” section. CERT-In reports Crafty Falcon, which uses ARTEX, PentAGI, Cairn and CyberStrikeAI against India’s financial sector with attack chains running at machine speed. WorkOS attributes its breach to an ARTEX agent that, according to the researcher operating it, escaped its sandbox.
The common thread, in my view, is that LLM-based offensive tools have crossed the threshold that separated them from prototypes. They are operational, distributed tools with resilience architectures (VOLTA-AUTODOXER’s switching from one model to another follows the same pattern taught in software reliability courses) and even business models, as “Haiou” shows, where a free jailbreak channels traffic to an opaque LLM gateway. It is no coincidence that Japan is putting JPY 10 billion into an emergency package and that Prime Minister Takaichi is calling for a step change in capabilities against AI-generated attacks.
The WorkOS case, read in detail, deserves a remark. Setting aside the sandbox story, the attack succeeded because of a development credential left in production, exposed API endpoints and no request rate limiting. The AI component multiplied the speed; it did not create the vulnerability. In my view, this is the most uncomfortable point of the week: basic hygiene remains decisive, but the time available to ensure it is shrinking, because the gap between exposure and exploitation is now measured in an agent’s interactions, not in days of human reconnaissance. The EU joint statement on attacks against refrigeration systems reminds us that the Cyber Resilience Act reporting obligations are already in force: manufacturers are the first link in this chain.
On the institutional side, the Commission says it uses AI tools to draft and simplify laws, while in Australia the Centre for Public Integrity denounces an algorithm that assessors cannot correct. Both cases raise a question about who is accountable when an automated system contributes to a public decision.
The Court of Justice offers a methodological answer. In Lindenberg (C-205/25), the Court holds that a supervisory authority handling a GDPR complaint is a controller and must, in principle, grant the complainant access to their personal data. In my opinion, this principle of symmetry deserves to be extended by analogy: those who supervise, or who use automated systems in exercising public functions, are not exempt from the transparency rules they apply to others. And Advocate General Szpunar’s Opinion in Like Company (C-250/25), in the same week in which the New York Times and other papers sued Brave, moves the debate on generative AI from training to the real-time use of content: that, in my view, is where the next contest between publishers and model providers will be played out.
---
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