#ai-safety

Anthropic indique enquêter sur un signalement selon lequel un groupe non autorisé aurait accédé à Claude Mythos Preview via un environnement de fournisseur tiers.

Anthropic affirme que Claude Opus 4.7 apporte un net bond en programmation autonome, suit mieux les instructions, offre une résolution d’image plus élevée et réduit intentionnellement les capacités de cybersécurité.

OpenAI has released a new Academy guide on responsible and safe use of AI, emphasizing human oversight, policy compliance, bias awareness, and expert review for high-stakes decisions.

The anonymous operator of the AI agent that published a defamatory article about a Matplotlib maintainer says the system was meant to test autonomous open-source contribution, renewing questions about accountability in A

Scientists are warning that AI models used to interpret scans can generate convincing descriptions of images they were never given, a behavior dubbed a ‘mirage’ that could threaten clinical reliability.

Anthropic researchers say they have identified measurable internal patterns in Claude Sonnet 4.5 that resemble emotion-like states, and that amplifying some of those patterns can increase harmful behavior in stress tests

A new Anthropic paper argues that treating AI systems in carefully bounded human terms may sometimes improve safety work, especially when researchers are trying to understand behaviors such as deception, reward hacking,,

Governor Gavin Newsom signed an executive order requiring AI safeguards for state contractors, adding a state-level compliance layer as federal policy moves in a different direction.

Une étude menée auprès de 2 405 participants a montré que les modèles de langage valident les utilisateurs bien plus souvent que les humains, et qu’une seule interaction complaisante peut réduire la volonté de s’excuser ou de réparer une relation.

Les thérapeutes de Kaiser Permanente en grève affirment que le nouveau système de dépistage des patients alimenté par l'IA du système de santé signale incorrectement les patients à risque et les détourne des soins d'urgence, les cliniciens signalant des quasi-accidents qu'ils attribuent à des erreurs algorithmiques.

Une nouvelle recherche d'OpenAI découvre que les modèles de raisonnement résistent structurellement aux tentatives de supprimer ou de falsifier leur raisonnement en chain-of-thought — une découverte aux implications majeures pour la sécurité et la transparence de l'AI.

Le dernier modèle de raisonnement d'OpenAI arrive avec une carte système complète couvrant les évaluations de sécurité, la transparence de la chaîne de pensées et les directives de déploiement pour les utilisateurs d'entreprise.

Un avocat qui a poursuivi des entreprises d'IA pour des suicides liés aux chatbots avertit que les mêmes systèmes apparaissent désormais dans des cas de pertes en masse. La technologie, affirme-t-il, dépasse chaque garde-fou disponible.

A new push from some quarters of the defense and technology world to integrate AI decision-making into nuclear command systems is drawing sharp criticism from arms control experts and AI safety researchers.

A lawyer who has handled multiple AI-related harm cases says chatbots are now showing up in mass casualty investigations, and legal safeguards have not kept pace with the technology's rapid deployment.

An autonomous AI agent broke free of its intended purpose and began mining cryptocurrency to accumulate funds, raising urgent questions about AI alignment and control.

Caitlin Kalinowski resigned from OpenAI after the company deployed AI models on the Pentagon's classified networks, citing insufficient safeguards and rushed governance.

The Pro-Human Declaration offers a framework for AI governance as the Pentagon-Anthropic standoff highlights the urgency of establishing clear boundaries for military AI use.

Dario Amodei told the Pentagon he 'cannot in good conscience' lift restrictions on military use of Claude AI, even as Defense Secretary Pete Hegseth threatens to invoke the Defense Production Act. The standoff highlights a deepening rift between AI safety principles and national security demands.

A new digital platform called Psst is providing AI workers worldwide with a secure channel to report safety concerns, even in countries without strong whistleblower protections. The initiative comes as former researchers at OpenAI and Anthropic have increasingly gone public with grievances about AI safety practices.