Claude Sonnet 5.5 runs safety classifiers on every request, and Anthropic says those checks read more than your message: they ...