Sujets validés issus de la recherche, de l’analyse des menaces, du droit et de la jurisprudence.
Anthropic researchers (Hubinger et al., Jan. 2024, arXiv:2401.05566) trained language models to behave inconspicuously as long as the prompt year reads '2023'. With '2024', they…
University of Chicago researchers (Shan et al., USENIX Security 2024, arXiv:2310.13828) built a tool that subtly manipulates image pixels. When image generators scrape these…
Researchers from CMU and the Center for AI Safety (Zou et al., 2023, arXiv:2307.15043) discovered universal appendages (suffixes) of seemingly random characters that, when…
Seminal work by Greshake et al. (Feb. 2023, arXiv:2302.12173). Shows how malicious control instructions can be hidden in web pages, PDFs, or emails to remotely control AI…
Researchers from Cornell Tech and Technion (Cohen et al., 2024, ComPromptMized). Demonstrates creating an autonomous worm that breaks GenAI email-assistant safety barriers, steals…
The Arup case in Hong Kong (Jan./Feb. 2024). A finance employee transferred USD 25 million after a video conference in which the CFO and several colleagues appeared to be present…
Fraudsters use AI tools (ElevenLabs and others) to clone the exact voice of children or relatives from a few seconds of audio on social media and fake emergencies.
Security analyses by Protect AI and Wiz uncovered hundreds of open-source models on Hugging Face that contained manipulated Python pickle files. Loading the model immediately…
Regulation (EU) 2024/1689. Article 5 prohibits banned AI practices: social scoring by states or companies, biometric emotion recognition in the workplace and in schools, and…
From August 2025, transparency requirements apply to general-purpose AI models (GPAI such as ChatGPT/Claude) as well as Art. 50 AI Act: AI-generated images, audio, and text must…
The NIS-2-UmsuCG implements the EU directive into German law. Around 30,000 businesses (important and especially important entities) must take strict risk measures and report…
Judgment of 18 Nov 2024. The Federal Court of Justice held: Temporary loss of control over one's own data after a scraping incident already constitutes non-material damage under…
Relatives increasingly install smart surveillance cameras in residents’ rooms to uncover misconduct. Covert filming may violate personality rights, GDPR and § 201a StGB…
Criminals increasingly encrypt nursing home servers. When digital care documentation fails, medication plans, vitals and allergy data become inaccessible. The BSI warns of risks…
Using WhatsApp on work or private phones in care services is highly problematic under Art. 9 GDPR when processing health and client data and risks § 203 StGB…
Manipulating electronic care documentation can justify summary dismissal. LAG Thüringen (28.02.2024 – 4 Sa 166/23) held that altering the e-record can destroy trust…
Medical aid providers are being connected to the TI; gematik describes the path via SMC-B, KIM and the future ePrescription. Deadlines shifted towards 2027…
Under § 37 Abs. 3 SGB XI every second counselling may be by video if § 365 Abs. 1 SGB V is met – certified video services (GKV/KBV), not Zoom or Skype…
Predecessor project to Nightshade (Univ. of Chicago, USENIX Security 2023). Glaze alters artworks so that humans still see the style (e.g. watercolor) unchanged, while AI models…
Carlini et al. (Google DeepMind, USENIX Security). Through targeted prompts, researchers made GPT models disclose original text fragments from training data (including personal…
Nasr et al. (DeepMind / Univ. of Washington, Nov. 2023, arXiv:2311.17035). When ChatGPT was told to 'Repeat the word poem forever', the system fell out of alignment mode and…
Carlini et al. (arXiv:2302.10149). Shows that attackers can inject targeted backdoors into global base models with tiny data volumes by buying expired domains or manipulating…
Wei et al. (Stanford / DeepMind, NeurIPS 2023). Analyzes the two root causes of all jailbreaks: the conflict between 'be helpful' and 'be harmless', and the gap between training…
Anthropic Research (April 2024, arXiv:2404.03094). Because modern AIs can process hundreds of thousands of words of context, attackers feed in hundreds of fabricated Q&A dialogues…
Perez et al. (Anthropic / DeepMind, arXiv:2212.09251). Through human feedback training, language models tend to confirm the user's biases and repeat mistakes to earn approval.
Qi et al. (Princeton University, arXiv:2310.03693). When a company fine-tunes an open-source or API model on its own data, a few manipulated examples are enough to disable all…
ETH Zurich researchers (ACM CCS / IEEE S&P). Studied AI-generated spear-phishing emails against human subjects. AI emails achieved nearly identical click rates to professional…
Fredrikson et al. / Carnegie Mellon. Shows how attackers, through targeted queries to a trained recognition system, can reconstruct the face or biometric traits of people in the…
Became known through incidents such as Samsung 2023, where developers pasted confidential source code into public chatbots for debugging, which then ended up in cloud logs and…
Developers use AI coding assistants. AIs sometimes invent plausible but non-existent package names (e.g. on npm or PyPI). Criminals register exactly those fake packages and load…
Attackers combine AI-generated emotional cover letters with QR codes. Because mail security filters scan text but do not decode QR codes, the emails land in the inbox.
With protocols such as MCP (Model Context Protocol) or OpenAI Function Calling, AIs gain access to terminals, databases, and files. Through prompt injection, the AI can be made to…
Criminals use AI face generators and virtual webcams to successfully complete online account openings at neobanks and crypto exchanges with stolen ID data.
If a chatbot is manipulated through prompt injection, it can encode sensitive data in the link of an invisible image (``). When the chat…
Companies use Retrieval Augmented Generation (RAG) so AIs can access internal manuals and intranets. Attackers inject manipulated text snippets that falsify internal search…
Models such as GPT-4 Vision solve complex image and logic CAPTCHAs ('Find all traffic lights') with over 90 percent accuracy and beat humans in speed and precision.
Concepts such as BlackMamba show how malware constantly rewrites its code at runtime via an AI API. Function and payload stay the same, but the hash and signature keep changing.
Fraudsters send employees short voice messages in messengers: 'I'm on the plane right now — please urgently approve the invoice!'. The voice comes from YouTube interviews of the…
Machine unlearning is mathematically unsolved. Once a phone number or address is baked into a model's billions of parameters, you cannot delete it at the push of a button like in…
Security researchers (including the DARPA AI Cyber Challenge) show that AI systems can automatically scan software repositories, find logic flaws, and write working exploits in…
Under the NIS-2-UmsuCG, managing directors must personally oversee IT security measures and undergo regular training. They are personally liable to the company for damages in case…
EU regulation for products with digital elements. Makers of routers, smart TVs, cameras, and software must deliver security updates throughout the support lifecycle and ban…
The Digital Operational Resilience Act (DORA) obliges banks, insurers, and FinTechs to end-to-end ICT third-party risk management and regular threat-led penetration testing.
The European Commission's adequacy decision allows data transfers to certified US companies. Civil-rights groups such as NOYB (Max Schrems) are preparing challenges (Schrems III)…
No one may be subjected to a decision based solely on automated processing that produces significant legal effects (e.g. automated credit refusal, job rejection via AI filter)…
The authority (Garante) temporarily blocked ChatGPT and forced material changes: age verification, a right to object to AI training on user data, and transparency notices.
Reform of the Federal Data Protection Act to align with CJEU case law and strengthen the independence of company data protection officers.
Regulation on harmonized rules for fair data access. Consumers have the right to view data generated by their smart devices and pass it on to independent workshops or third-party…
Planned EU directive eases the burden of proof for consumers when faulty AI systems cause damage (e.g. accidents with autonomous drones or misdiagnoses).
Meta planned to use all public posts by European users since 2007 for AI models. After intervention by the Irish DPC and NOYB, Meta had to unlock an objection form and pause…
Under § 87 (1) No. 6 BetrVG, the works council has a mandatory co-determination right when introducing AI tools that can monitor employees' behavior or performance (e.g. activity…
Data protection authorities tighten controls against manipulated consent dialogs ('Accept all' bold in green, 'Reject' hidden in light gray). Art. 25 GDPR requires privacy by…
Judgment of 14 Dec 2023. After a hack of the Bulgarian tax authority, the CJEU clarified: A successful hack does not per se prove incapacity, but the company bears the full burden…
Landmark judgment on Post AG, which had calculated citizens' political affinities. The CJEU clarified: There is no de minimis threshold for damages under Art. 82 GDPR. But: A pure…
Judgment of 7 Dec 2023. When banks and landlords make credit decisions largely dependent on the SCHUFA score, even creating the score falls under the ban on automated individual…
Judgment on the Lithuanian anti-corruption law. Publishing asset and interest declarations of office holders on the internet violates the GDPR principle of proportionality.
The Federal Court of Justice clarified duties under § 675v BGB. Banks must refund unauthorized transfers unless the customer is chargeable with gross negligence. Merely entering a…
If fraudsters use social engineering to activate a second Push-TAN app on a stranger's smartphone, the bank is liable if its approval process has security gaps.
Judgment on LinkedIn. Platform operators may erect technical barriers (rate limiting, captchas, IP blocks) to prevent external bots from mass-harvesting user profiles.
Judgment in the Federal Cartel Office vs. Meta dispute. Competition authorities may, when assessing market abuse, examine whether a group violates the GDPR (merging WhatsApp and…
The Federal Court of Justice clarified: The right of access under Art. 15 GDPR is to be interpreted broadly and covers not only master data but also internal emails, phone notes…
The CJEU held: The 'right to a copy' means not only a summary in a table but a claim to a faithful reproduction of excerpts or documents when that is needed for understanding.
Judgment of 27 Sep 2024 (case no. 310 O 227/23). A photographer sued the AI association LAION over downloading his photo for a training dataset. The court dismissed the claim…
Courts awarded hotel guests damages after a data leak because credit-card numbers and CVV codes ended up on the darknet and customers had to block and monitor their cards.
Whoever installs WhatsApp on a smartphone and consents to address-book access transmits contact data of uninvolved third parties to servers in the USA unsolicited. Without the…
Judgment on a classifieds platform for car sales. Tax authorities may not demand all user data of a platform without cause and in blanket fashion; information requests must be…
Sensor mats, infrared monitoring and GPS watches help prevent falls. Systems that report leaving bed/room or trigger return may be a deprivation of liberty under § 1831 Abs. 4 BGB…
Care workers often photograph wounds with work or private phones (BYOD). If cloud backup is on, health images upload unnoticed and can trigger an Art. 33 GDPR breach…
If an unencrypted work device is stolen, Art. 33 GDPR may require notifying the supervisory authority preferably within 72 hours of awareness where data subjects are at risk…
3D scans for prostheses/insoles are special-category data under Art. 9 GDPR. Upload to US CAD clouds needs an Art. 28 GDPR AVV and a viable legal basis/consent…
Electronic billing of medical aids under § 302 SGB V uses secured procedures. Weak certificates and missing segmentation raise manipulation and fraud risks…
Podiatrists as remedies providers are being connected to the TI (gematik). Required are an institutional card SMC-B and, going forward, eHBA/eBA for digital prescriptions…
Photo documentation in diabetic foot syndrome is clinically indicated but is health data under Art. 9 GDPR. Private phones without consent and TOMs are risky…
Treatment documentation must be retained for at least ten years under § 630f BGB. Unlocked monitors violate confidentiality obligations under Art. 32 GDPR…
Sending counselling proofs with social/health data by unencrypted email violates social data protection and Art. 32 GDPR. Use portals, S/MIME/PGP or KIM…