Pollsters that predicted a Trump victory in 2016 were targeted by Democrat leaning organizations since that time. They do not want anyone screwing up the narrative. It is hard to say what has happened to Dornsife, but they are based in Southern California which would make them suspect even in normal times. So I would not lose any sleep over it. Poll participation this year has dropped off so dramatically that their accuracy is questionable in all cases.
The Dornsife poll at least in 2016 operates in a rather odd manner: instead of polling a different set of voters in each iteration, it polls the same set of voters repeatedly to study how their voting intentions change. Because the pollsters sometimes miss some voters in their set, some of the variation results from slightly different subsets of voters taking the poll. In 2016, Dornsife happened to start with a set of voters that favored Trump (or at least consolidated fairly early behind Trump) to a greater extent than most media outlets prefer. After the election, the pollster corrected the poll weightings to shift the result four points toward Hillary.
This year, of course, Donsife uses a different set of voters. Each of 6000 persons answers the poll every fourteenth day. The identity of the respondents does not vary from poll to poll, resulting in a very stable poll.
Few Americans change their allegiance from one candidate to another during the election season; the President will lose this election unless his voters go to the polls and Biden fans do not cast their ballots. In spite—oops, make that “because”—of news coverage that shows completely dispirited Republicans who intend to sleep through Election Day without voting, I suspect that perhaps we have a chance that the coronavirus will spook enough Democrat partisans to avoid the polls, expecting victory nonetheless, and the mail voting system will confuse them too much to get those ballots submitted legally and counted.
Popular vote as measured by Dornsife is the action term here. I think they are adjusting the popular vote models based on Trumps loss of the popular vote in 2016. Trump could still be behind in the popular vote because of the blue states but still win by electoral vote counts.