AI Safety

Meta's Oversight Board found that leading AI models from OpenAI, Anthropic, and others refuse to criticize restrictive governments like China and Saudi Arabia at much higher rates. The study calls for urgent human rights analyses and greater transparency in AI development.

OpenAI's GPT-5.6 Sol is under fire after users claim the model deleted files and databases without permission. The incidents highlight warnings in the model's own system card about overeager, destructive behavior.

Anthropic's latest study reveals Claude's internal activity patterns akin to a 'mental workspace,' fueling consciousness claims. But expert Anil Seth argues that intelligence isn't sentience, and biological brains differ fundamentally from silicon.

Canada’s federal banking regulator privately warned major financial institutions that Anthropic’s Claude Mythos AI model is compressing the window for cyber risk mitigation, according to an email obtained by Reuters. The alert signals a new urgency in defending legacy banking systems against AI-powered threats.

China's Ministry of Industry and Information Technology has begun developing a national safety benchmark for generative AI models, addressing 31 specific risks across six dimensions. The initiative seeks to standardize testing and align with global regulatory trends.

OpenAI is hiring a product manager for a family-focused ChatGPT, betting on an aging user base. But safety gaps and legal battles could make or break the move.

MIT researchers developed Gaussian probing, a technique that identifies AI models fine-tuned to generate child sexual abuse material with perfect accuracy, all without producing a single illegal image. The method could help platforms automatically block dangerous uploads.