The last month, Anthropic: - Quietly nerfed their flagship model harness (Claude Code) without telling anyone - Banned corporate customers of Claude - Silently changed plans for customers with certain files in their repo All evidence that closed models are *massive* risks.