Why does the computer think black is winning?

I don't understand the evaluation. It's just a draw, but the computer gives almost a -12 advantage to black.

Originally from a puzzle in the "Winning Chess Tactics" book from Yasser Seirawan (Test 90, page 122). In that puzzle the position is lost, although this is given as an example of white trying to defend and not lose! You can simply avoid taking the bishop after 1....Ba7+ (taking either with the rook or the king and then 2....b6 leads to a draw).

Any ideas?

Why is this modified position a draw? If there was no Ra8 I would agree but why black cannot move the Bc8 away, move the king out of the way of the rook, bring the rook to the g file, then win easily? 

Because you have time to get your king to g1 and then h1. From there the black pieces have no way of infiltrating your position, even though black is up two pieces (example moves below).

As you can see here the black bishop is useless, and the rook can't do much (Rook to g2 is stalemate, to g1+ just gets captured, and any other move just doesn't let black break through). But the computer still evaluates a huge advantage for black. Weird.

OK I see.. I guess the depth of analysis is not high enough (like mine lol). 

Its not a depth of analysis issue.  Stockfish can get to depth 58 pretty quickly.  Its that there are so many possible moves that don't repeat x3, that stockfish still thinks its winning and keeps trying anything that doesn't lead to a repetition.  Its the kind of position that is designed to flumox engines.  I'd be interested to see if Leela does better.

Fortress draws are not evaluated as 0.00 by traditional engines. Instead they have a constant evaluation that won't change if you shuffles the pieces a little bit. My Komodo 10 gives a permanent -11.95 after 1.Ba7 Ra7 2.b6 Ra8 3.Kc3. In human words, Black has a huge material advantage that can't be converted into a win.

Another example of constant evaluations for fortress positions:

This is obviously a draw; Black simply shuffles his bishop along the f1-a6 diagonal. My Stockfish gives +5.07 at depth 30. It doesn't change as the depth approaches 60, indicating a fortress draw.

I just had something like this.  Down two pawns, I managed to salvage a draw like so (I'm playing White):

At least I'm pretty sure that this is a draw from at least after Black's 48th move, despite the engine giving -4.60 at the point where we virtually shook hands.  I'm not missing something, am I?

To OP: Most engines don't evaluate the real chances of increasing the activity to a point of becoming unstoppable. Instead, they use "positional" criteria when their calculation horizon doesn't reach a conclusive result.

In those examples, Black is material up and there's no compensation for White, so the engine "assumes" that there "must exist" a method to convert the advantage but that it hasn't found it yet. That said, given the resources and time, the engine is capable of realizing that there's no forceful way to win such positions within the 50-move rule and give a 0.00 value.

The modified position is not a draw I think. Bishop F5 I think wins for black. If you capture you lose a tempo and if you don't the rook can reach the g file in time to cut off

I think OP mean: White to move and draw!

In the original position (#1) If it's White to move it's stalemate