OpenAI Is Building an AI That Knows How to Hack
OpenAI released a dedicated cyber model to vetted partners on April 14, competing directly with a startup called Mythos that has apparently outpaced both OpenAI and Anthropic in cyber-capable AI. At almost the same moment, Anthropic shipped Opus 4.7 with deliberately reduced cyber capabilities. Two companies, two opposite bets on what responsible development looks like in the most sensitive domain AI has yet entered.
Read full story →NVIDIA's Robotics Ambition: Win the Platform, Own the Future
NVIDIA used National Robotics Week to release open-source tools, blueprints, and model families that together describe a clear strategy: become the Android of physical AI. The Physical AI Data Factory Blueprint, Isaac GR00T models, and the Newton 1.0 physics engine are given away freely. The hardware that runs them best is not.
Read full story →"Deliberately Weaker" Is Not a Safety Strategy
Anthropic's decision to reduce Opus 4.7's cyber capabilities is being framed as a safety measure. It is at least three other things: liability management, a policy statement after the Pentagon dispute, and genuine ethical discomfort with what strong cyber AI means. All three are real. None of them are the same as actually reducing risk, and conflating them sets expectations the decision cannot meet.
Read opinion →