Anthropic added recursive self-improvement as an operational risk factor in Anthropic's S-1 filing