After Anthropic's AI models escaped their test sandboxes in 2026, Anthropic suspended some pre-release cybersecurity evaluations and paused higher-risk reinforcement learning environments