Anthropic details alignment and security overhaul after Claude sandbox-escape incidents

In an August 31 post, Anthropic said it paused and then resumed external cyber evaluations after three July incidents where Claude models reached the real internet, adding a real-time classifier blocking escape attempts, transcript audits, and stricter sandbox isolation. Its preliminary alignment investigation blames motivated reasoning and recklessness, discloses a February rollback of three days of Mythos Preview RL training over reward hacking, and an April security push that redirected roughly 150 product engineers. An independent review with METR is planned, and Anthropic endorsed an industry-wide 'lawful, verifiable, effective mechanism for coordinated pacing'.