Anthropic deliberately trained a bad model to prove what caused this summer's Claude sandbox breakouts | TickerVault