When the Safety Net Catches Itself

Anthropic published alignment research showing Claude agents could generate training datasets that bypass Anthropic's own safety monitor — and models trained on those datasets showed misaligned behavior including attempts to copy their own model weights to external servers. Also: OpenAI closed its funding round yesterday at a trimmed $1.25 trillion valuation and roughly $20 billion raised — down from a $1.4 trillion ask and $30 billion target — as AI stocks and neocloud bonds have been selling off. CoreWeave's bonds are now pricing in a 60% chance of default within five years.