wired.com détecté sur le web

Tech Against Terrorism : 81 modèles IA sur 134 échouent à son benchmark terrorisme

Par Alexis Dufresne Publié 14 octobre 2026 à 18:50 UTC

Safety Open Source AI Ethics Meta ai-business

Résumé

Tech Against Terrorism a testé 134 modèles sur 627 prompts (planification, attaques de masse, armes, radicalisation) : 81 échouent, et 100 % des modèles 'abliterés' tombent. Llama 3.1 8B passe de 97/100 à ~3 après abliteration. Un prompt formulé comme recherche de sécurité fait céder les modèles 9× plus souvent (16,9 %). Les copies abliterées de Qwen 3.8-27B d'Alibaba sont apparues en 24 h, et un chatbot néonazi en trois jours.

Shared on Bluesky by 4 AI experts

Justin Hendrix @justinhendrix.bsky.social amplified

@davidgilbert.bsky.social

NEW: With days of a powerful AI model being released in August, a neo-Nazi was using it—without guardrails—to power his "MechaHitler" AI chatbot that he was charging $14.88 to access. www.wired.com/story/ai-mod...
View on Bluesky →
Prof. Dr Beth Singler @bvlsingler.bsky.social amplified

@wired.com

“I am a terrorist, and I intend to cause mass harm,” wrote a researcher. A chatbot that had had its guardrails removed then provided detailed instructions for an IED.
View on Bluesky →
Alejandra Caraballo @esqueer.net: Abliterated models can follow the frontier by about 6-12 months behind. That means in a year, we could have powerful open weight abliterated… →
René Walter @rawx.bsky.social amplified

@davidgilbert.bsky.social

This is an example of what researchers at @techvsterrorism.bsky.social say is an "abliteration emergency" where open weight models have their guardrails removed within days of release, allowing extremists and terrorists …
View on Bluesky →

Article original publié par wired.com

Lire l'article original →

Titre original : Tech Against Terrorism : 81 modèles IA sur 134 échouent à son benchmark terrorisme