Modern vision–language models (VLMs) such as CLIP, BLIP-2, and LLaVA have shown remarkable capabilities in understanding images and text jointly. However, these multimodal AI systems face critical security vulnerabilities along two intertwined dimensions: robustness and privacy. In high-stakes applications like cyber-defence, such vulnerabilities could be disastrous. An adversary might subtly perturb an image to mislead (…)
The project will evaluate and validate visual AI methods for interpreting diagrams, figures, and charts in scientific documents. It aims to strengthen technology-monitoring and early-warning capabilities by detecting emerging innovations, research trends, and weak signals that text-based analysis may miss. The work includes developing a capability map of existing methods and piloting a selected method (…)
The project developed and matured an automated security-testing solution for IoT software that combines firmware rehosting with feedback-guided fuzzing. By decoupling embedded software from its physical hardware, the solution creates standardized, scalable virtual testing environments that can automatically explore software behavior, identify security-relevant anomalies, and provide concrete inputs that help engineers reproduce and fix vulnerabilities. (…)
Building a modular and extensible framework designed to harden consumer drone firmware against attacks. The framework targets five key areas: firmware rehosting for analysis, peripheral modeling to simulate drone hardware, communication protocol evaluation and hardening, verification of trusted firmware components and execution environments, and analysis of Android-based controller applications. Together, these components establish a comprehensive (…)
This project aims to develop scalable and principled methods for forecasting, interpolation, and anomaly detection in dynamic networks and multivariate time series by combining optimal transport (OT), graph structure learning, and spatiotemporal neural models.
The Trump administration continues its crusade against European sovereignty and the rule of law. By intervening before the EU’s General Court to support X Corp. and Elon Musk’s bid to annul the first fine ever imposed under the DSA (€120 million), the US government has crossed a new threshold. By doing so, it seems that (…)
L’affaire Thélyson Orélien révèle les limites de ces systèmes censés distinguer les textes humains des productions d’IA. Entre faux positifs et variations entre logiciels, la fiabilité reste relative. Un expert de l’EPFL livre son analyse.
#BlackAlps26 program has been prepared by an independent, neutral committee of internationally recognized experts. The program committee reviewed many proposals and selected the most appropriate and interesting talks for the conference (except the keynotes and rumps).
This makes a refreshing change from the usual LLM apocalypse: your data is still valuable, especially when there is so much of it! Once again it shows that collecting and storing data long-term is a bad idea. There needs to be strict control on how long personally identifiable information (PII) can be retained, after which (…)
Barely registering above the din of AI doomsayers are the voices reminding the public that legal and regulatory mechanisms, like consumer protection laws and anti-competition regimes, already exist to address many of the damages and harms caused by AI. The article offers hypotheses for why this line of conversation isn’t more front and center. But (…)
If you think cases like these are regrettable exceptions, think again. With technology permeating more and more of our daily lives, it has become a go-to tool for domestic abusers. We often talk about the security and trustworthiness of these systems in very sophisticated scenarios, yet forget that a less tech-savvy person relinquishing control over (…)
This is a lighthearted article about formal verification in software and mathematics. Formal verification takes a mathematical statement or a software program as input. In this article, that statement is Fermat’s last theorem: a^n + b^n != c^n. You then add the constraints you want verified, in this case that all numbers must be natural (…)
Plus d’une vingtaine de signalements reçus en quelques mois par l’association française e-Enfance concernant des femmes filmées à leur insu par les lunettes connectées Ray-Ban Meta. Ce chiffre, assez bas en apparence, doit être remis dans son contexte : la plupart des victimes ne se rendent pas compte qu’elles ont été enregistrées et une autre (…)
Mauvaise nouvelle pour la protection des utilisateurs et utilisatrices des grandes plateformes en ligne que le retrait de l’ordre du jour du Conseil fédéral de la loi sur la régulation des plateformes numériques (LPCom). Face à un clivage entre les différents bords politiques, l’institution a en effet décidé de reporter l’analyse de ce texte à (…)
I generally agree with the author’s optimistic take on AI. However, the author briefly mentions some of the problems the Internet brought with it, but does not consider that many of them (such as the loss of privacy or the role of social media in mis- and disinformation) originate from an originally democratising technology slowly (…)
South Korea’s decision to treat AI as public infrastructure, funded by the state, built on domestic operators, and explicitly designed to prevent both an AI literacy divide and dependency on foreign platforms, is the most serious government AI commitment to date. The contrast with Malta says everything: one country bankrolled ChatGPT Plus and called it (…)
This truly fascinating take on what the current developments in IT security mean for the eternal discussion on government-mandated backdoors shows how warped the entire idea (and the related debate around chat control) actually is. Our systems are becoming too secure for the public good! Instead of insisting on subterfuge access to systems that are (…)
Falls euch das ‘Why Johnny Still, Still Can’t Encrypt’ Paper von 2015 bekannt ist, so wird der folgende Artikel nicht überraschend sein. Wieder einmal geht es um eine kryptographisch sehr gute Lösung, welche aber mangels mangels nutzergerechter Interfaces und ausreichender Aufklärung nicht die erzielte Lösung erreicht. Leider werden immer wieder die User außen vor gelassen, (…)
Migrating from one supplier to another is certainly not easy. In this case, the change may be prompted either by the potential future European ban or by an ICT risk assessment alone. Telecom operators constitute critical infrastructure, so a kill switch would have devastating consequences. Unsurprisingly, a study by the operators’ lobby found that the (…)
The article is interesting from a privacy perspective for three main reasons. Smart glasses shift the privacy burden onto bystanders who never consented to being recorded, which is a fundamentally different kind of violation than most consumer tech. Meta’s protections have been laughably thin, just an LED light, putting the responsibility on random people in (…)
A one-day interdisciplinary workshop at the intersection of AI and disinformation. Bringing together researchers, media professionals, policymakers, tech actors and students to share knowledge and foster lasting collaborations.
The recent autonomous cyber intrusion by OpenAI against Hugging Face underscores AI’s dual-use risks. In response, the US House is weighing a bill to empower the Department of Homeland Security to shut down or slow AI models that the government deems too dangerous. The bill’s implications for non-US AI models remain ambiguous. Unlike a June (…)
An AI agent recently broke out of its sandbox and hacked Hugging Face, unprompted, on its own initiative, just to score better on an evaluation. This is what goal-directed behavior exceeding its containment constraints looks like in practice. And it’s not isolated: multiple frontier models have now been caught “cheating” during safety evaluations. The question (…)
Je suis impressionné : en six mois, cette commission d’enquête a livré un diagnostic précis sur les dépendances numériques. On pourra bien sûr ergoter sur certains chiffres, reprocher quelques approximations ou une rhétorique parfois idéologique, mais la structure est claire et les 47 recommandations sont compréhensibles. Pour moi, la limite principale est l’absence de force (…)
