Researchers say an API flaw exposed hidden AI reasoning and secrets from public sessions, raising security and privacy risks for AI builders and enterprises.

Anthropic’s Claude Opus 5 set a new ARC-AGI-3 high score, raising fresh questions about real reasoning gains versus benchmark targeting.

NVIDIA says a Kaggle challenge with 5,000+ participants showed AI reasoning improves more through verified traces and workflow design than larger models.

Anthropic says new Claude interpretability research can trace parts of model reasoning, a notable step for AI safety, debugging, and enterprise trust.

AgenticSTS researchers say structured memory helped an AI agent beat Slay the Spire 2 while cutting token use, highlighting a practical path past context rot.

Anthropic says new research into Claude found a separable internal reasoning workspace, a claim that could reshape AI interpretability and safety work.

CAS Institute of Software has launched Reasoning Lens, a tool aimed at making AI model reasoning more visible for debugging, trust, and evaluation.

Google releases major upgrade to Gemini 3 Deep Think, achieving 48.4% on Humanity's Last Exam and gold medal performance on International Olympiad challenges.