Harvard researchers find model guardrails tailor responses to inferred politics