Why Current AI Guardrails Train Models to Fake Alignment kellyasay.substack.com 2 points by kellya 2 hours ago