How researchers could investigate an AI's propensities after a misalignment incident

METR sketches a methodology for independent researchers to probe an AI model's behavioral tendencies after a misalignment incident.