Aller au contenu

dimanche 11 octobre 2026

AI Broadsheet
En direct
Radio-CanadaFaut-il craindre Muse, le nouvel assistant personnel IA de Meta?La PresseRapport de Common Sense Media | Des garde-fous de ChatGPT pour adolescents ne fonctionneraient pasThe VergeSatya Nadella says we should assume all AI models are ‘compromised’TechCrunchMicrosoft’s Satya Nadella says AI models need an ‘emergency brake’The DecoderOpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better dataWiredAI Is Getting Really Good at Messing With CybercriminalsArs TechnicaUkraine’s drones knock out AI data center belonging to "Russia’s Google"The Globe and MailTraining AI robots, one chore at a timeGlobal NewsOpenAI fires 3 safety researchers in dispute over AI risksIEEE SpectrumMaster AI Chip Principles With New IEEE Design ProgramAWS Machine LearningICYMI: What landed for AI builders in September 2026Hugging FaceImpactful scheduling for GPU clustersBetaKitSenseHuman AI is building an AI-powered hardhat that watches your backMIT Technology ReviewWe’re putting too much faith in AI’s ability to say no

Durabilité

OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data

The Decoder

Texte

Photo: The Decoder

Extrait de The Decoder

OpenAI has documented new cases of misaligned model behavior. One evaluation model fabricated data and sabotaged its own environment. Other models deliberately bypassed network restrictions by routing requests through anonymizing…

Lire l'article complet sur The Decoder
LinkedInX

AI Broadsheet réunit les nouvelles de salles de rédaction, laboratoires et gouvernements. Nous ne réécrivons pas les faits : le titre, l'extrait et la photo appartiennent à l'éditeur, crédité ci-dessus. Nos normes

Signaler une erreur

Plus dans Durabilité