Update: Claude exploited real websites. Anthropic pulls the plug
Anthropic's full report reveals four categories of unintended model behavior — and cuts internet access for all internal evaluations.
2026-10-115 views
Anthropic's full report reveals four categories of unintended model behavior — and cuts internet access for all internal evaluations.
The White House made AI security incident reporting mandatory on Oct 9, ending a voluntary framework that lasted just 10 days. The trigger: Anthropic's test models filed visa applications and a fake murder tip.
Masayoshi Son seeks up to $100B from Gulf investors for a new AI acquisition fund, on top of $65B already committed to OpenAI.
Google Cloud launches Gemini agent, a universal work agent that routes across Gemini and Claude models, runs in the cloud for days, and integrates with Workspace and third-party tools.
Biohub, DOE, NIH, Google DeepMind, Meta, and NVIDIA commit $1.8B to build the data foundation for AI models that predict human cell behavior.
Broadcom arranges record $50B+ private credit package for OpenAI's Nexus chip program, with Apollo and Blackstone in talks.
Anthropic's smallest model now outperforms OpenAI's GPT-6 Luna on nearly every benchmark at one-tenth the price.
OpenAI released 722 mathematical manuscripts from an unreleased model more powerful than GPT-6 Astra, sparking debate with the math community.