This is pure speculation. You can't see into its mind. Commercially implemented AIs have recommended recipes that involve poison in the past, including one for mustard gas, so to give it the benefit of the doubt and assume it was even tangentially correct is giving it more slack than it has earned.