Anthropic retrained Fable 5's safety classifier to stop rerouting benign biology questions -- an 85% cut in fallbacks -- while keeping virology, toxicology and molecular design blocked, days after Stanford published AI-designed viruses and hours after OpenAI paused a model over cyber risk.

The IWF and NSPCC's Report Remove service took 420 reports from UK children of manipulated explicit images of themselves in six months -- already past its 397 for all of 2025, with most severe enough to meet the criminal threshold.

Climate scientist Zeke Hausfather logged eight weeks of his own Claude Code use -- 1,138 prompts, 14,000 model calls, 3.2 billion tokens, about 170 kWh -- roughly 150 Wh a prompt against Google's published 0.24 Wh median, though the conversion to electricity is his estimate with a 70-330 kWh range.