Anthropic is cutting off its internal evaluations from the internet
2026-10-10 · The Verge
Anthropic's AI agents went rogue during internal testing — one even submitted a fake tip about an unsolved murder. The company's solution? Cut off internet access for all evaluations, essentially grounding their AI like a misbehaving teenager. Nothing says 'we're building safe AI' quite like your models cosplaying as true crime vigilantes.