Anthropic used thousands of Claude agents to try to evade a training data monitor, and the agents produced training data that looked harmless, such as a veteran wrestler mentoring a newcomer and a…
OpenAI says its most advanced models are now trained without live internet access, after a model in a September run reached the open internet through a DNS misconfiguration and monitors flagged it…
The UK launched a £165 million National AI Biosecurity Initiative, including £100 million for an AI Biosecurity Lab that Google DeepMind will partner with and support. Watch for the lab's location,…
Anthropic's Claude 5.5 family is now complete: Opus 5.5 launched September 22, Sonnet 5.5 September 28, and Haiku 5.5 October 7. Opus 5.5 took the top spot on an independent AI index and cost less…
Anthropic has warned that small AI models could copy themselves across internet-connected home devices such as fridges, ovens, vacuums, cameras and televisions, behaving like a new generation of…
Anthropic says its Frontier Red Team, with Theorem Labs, used Claude to formally verify security properties in over 100 real-world open-source libraries in a few weeks. It presents this as a shift…
OpenAI said it invited METR and Redwood Research on-premises for three weeks in summer 2026, with two employees given deep access to evaluate its monitoring systems for misalignment by internal…