Lê Nguyên Hoang

Lê Nguyên Hoang

@science4all.org

AI Security Assistant Professor @TélécomParis. @Polytechnique X07, @polymtl PhD, ex @mit @epfl. Writer, @Orange AI ethics, Science4All.

7.4K followers337 following422 posts

Les risques cyber gagneraient à être un argument plus systématique contre la numérisation accélérée de nos organisations. Surtout tant que les standards de cybersécurité sont catastrophiques et que les fournisseurs d'IA agentiques arment les cybercriminels. www.rts.ch/info/monde/2...

La communauté scientifique semble unanime : il faut être alarmiste vis-à-vis de la vulnérabilité de notre espace numérique, suite au déploiement des dernières IA. Des revendications drastiques ont été proposées dans cette lettre co-signée par 483 académiques (moi compris ^^). ai-security-letter.org

Je recommande inconditionnellement le film The Social Reckoning sur l'investigation sur les réseaux sociaux la plus compromettante de l'Histoire, actuellement au cinéma. Très juste, drôle et vibrant, c'est un film d'intérêt général pour la survie de nos démocraties. en.wikipedia.org/wiki/The_Soc...

The discourse around "AI misalignment" is dangerously offloading AI directors from liability. But US Treasury chief wants them legally responsible for the damages that their products caused. It is high time that actual AI security be a legal prerequisite. www.theregister.com/security/202...

theregister.comTreasury chief says AI bosses, not their bots, will carry the can for criminal acts'Humans are responsible, not the AI,' argues Scott Bessent as he calls out OpenAI agents' hack of Hugging Face

Heureux de vous announcer que je débute un poste de Maître de Conférence en Sécurité de l'IA à Télécom Paris !! Je suis impatient de poursuivre ma recherche dans le domaine, et de donner des nouveaux cours sur le sujet aux étudiants, qui consolideront nos connaissances scientifiques à ce sujet.

View full thread

Lê Nguyên Hoang reposted

*until 2026. I started this thread back in February, and the judicial progress to enforce long-lasting laws to recommendation AIs has been astounding. The trick? We're finally considering that algorithmically selecting published content is NOT merely hosting it. mu.social/profile/scie...

Lê Nguyên Hoang@science4all.org

Wooo !! Monumental. Parler de potentielle plus grosse amende de l'histoire ne fait pas justice au montant ahurissant exigé. 1400 milliards de dollars, c'est 42% du PIB de la France. Un tribunal californien a maintenant la survie de Meta entre ces mains. siecledigital.fr/2026/07/07/p...

View full thread

Of course, what actually matters is the greenhouse effect itself, and there's still a very long way to go. But isn't it motivating that the trend of the trend of the trend of greenhouse effects is clearly in the right direction, and that we simply need to further accelerate it?

What seems at least pretty clear by now is that, thanks to regulations and transition towards renewable energies, the third derivative has been negative for the last decade. Put differently, we are pushing to push the faucet towards fewer yearly emissions!

Lê Nguyên Hoang reposted

Lê Nguyên Hoang reposted

Counter doom by: 1. Communicating the threat accurately, 2. prebunk against emotionally manipulative narratives, & 3. pair the problem with credible solutions & agency. Full study & coverage below: doi.org/10.1073/pnas...

Sander van der Linden@profsanderlinden.bsky.social

Climate doom may discourage climate action, but here's a way to counter it, our new research in @theconversation.com with the brill @racheloboyle.bsky.social @tylershores.bsky.social! theconversation.com/climate-doom...

Lê Nguyên Hoang reposted

Lê Nguyên Hoang reposted

Meta is on trial. 29 U.S. states allege it designed Facebook and Instagram to keep young people hooked while misleading the public about the risks to children. This case could help hold Mark Zuckerberg’s company accountable for putting profit ahead of kids’ safety.

Lê Nguyên Hoang reposted

Lê Nguyên Hoang reposted

Lê Nguyên Hoang reposted

NEW: Meta has run dozens of paid ads that include explicit, AI-generated child sexual abuse material in recent months. Some as recently as this week. "These are ads that were reviewed, approved, and allowed to run by Meta, never encountering interference while the company collected the ad dollars."

wired.comMeta Ran Ads That Contained AI-Generated Child Sexual Abuse ImageryMore than 50 offending image and video ads were published across Facebook, Instagram, Messenger, or Threads, according to Meta’s ad library data. Some ran as recently as this week.

"Plus un échec de défense qu'une attaque exceptionnelle." Attention à ne pas sombrer dans un fatalisme contre les cyber-attaques des IA agentiques. Oui, elles peuvent détruire des châteaux de carte. Mais on peut aussi investir beaucoup plus pour construire de véritables forteresses numériques.

Lorenzo Franceschi-Bicchierai@lorenzofb.bsky.social

NEW: OpenAI’s hack against Hugging Face was novel because it was fully autonomous and AI-powered, but the rogue agent acted mostly human-like. And Hugging Face could have done a better job at spotting and stopping the attack with better traditional cybersecurity defenses, experts explained.

Une préoccupation récurrente, c'est de voir les entreprises d'IA privilégier la performance, quitte à faire chuter les standards de sécurité. Typiquement, des activités externes, massives et sans justification semblent devenues largement tolérées. Et oui, clairement, ça complique la cyberdéfense !

View full thread

Le Kospi, l'indice boursier des grandes entreprises coréennes, a longtemps explosé, porté par les valorisations des IA. Mais il chute depuis 1 mois. Éclater la bulle ne fera pas disparaître les IA, mais cela changerait certainement les investissements et la régulation. www.bbc.com/news/article...

Lê Nguyên Hoang reposted

New: Hugging Face has a widespread problem with nonconsensual deepfakes, including multiple AI models that undress or create sexual images. Researchers even created their own tool on the platform and tracked 1,000 prompts over a week. 73% of the prompts they received were sexual in nature.

wired.comHugging Face Has a Deepfake Nudes ProblemResearchers tested top image editing models on Hugging Face and found they could easily create explicit deepfakes—and 1,000 image editing prompts show how people use the software.

J'ai trop souvent vu des médias et des ONG inviter leurs employés pour parler d'IA, voire d'éthique de l'IA. Donnons à l'inverse les moyens financiers et légaux aux chercheurs indépendants d'effectuer des audits rigoureux, et de construire un numérique alternatif, sécurisé par conception.

Lê Nguyên Hoang reposted

alcoholic trying to get sober after a relapse tells Google's "Health Coach" about it. it tells them to drink every hour and messages them to make sure they're drinking as soon as they wake up.