It would seem like it's impossible for a commercial engine to make an obvious blunder that an average human player would spot. There are plenty of cases of free programs glitching. For example, one programmer accidentally made his engine evaluate longer mates as better than shorter mates, and it ran out of time after having mate in one each move when it was two Queens up for a few dozen moves. Also, Tim Krabbe has discussed a bug in Fritz 6 that evidently prevents it from seeing the idea of responding a check with a capturing Knight promotion in analysis (you can read about it in entry 152 at http://www.xs4all.nl/~timkr/chess2/diary_8.htm).

However, there is actually one unusual weakness that can apparently lead programs such as Fritz and Rybka to ruin simple endgames, assuming they are equipped with endgame tablebases. An endgame tablebase is a list of all positions up to a specific number of pieces including Kings (for engines it's usually five, although they have been calculated to six) and their evaluations with best play. This means that engines can play almost any endgame perfectly. There is, however, a strange exception. Consider this position:

What would you play as Black here? A reasonably strong human being would quickly play Qxh1+ K(any) Qxh2+ and exchange Queens, with a draw. However, endgame tablebases have a flaw: they don't include castling rights. An engine equipped with a tablebase will play Qxc2, transposing to a won endgame by tablebase, and when White plays 0-0+ the engine will suddenly realize it is lost. Of course, there is almost no chance that this would happen in an actual game, and I apologize if you consider this a waste of your time, but I find it quite interesting.

I imagine that once a tablebase-covered endgame has been reached in a real game, it's probably safe to assume that castling is no longer possible. Interesting observation though!

Interesting.

hahaha, shredder tablebase fell for that one.  even though the fen clearly states castling is still allowed.

Here's another possible situation. Picture someone playing Black against an engine. He gets into a difficult endgame, but he manages to escape into a R+B vs R endgame that is theoretically drawn, but very difficult. Black is subjected to this torture for a long time, and so far has managed to hold the draw. White has played 46 moves and Black has played 45 moves since the last capture. It's Black's turn.

EDIT: Oops, hold on, I'll insert the diagram in a moment EDIT2: OK, got it

Interesting stuff, SirDavid. Wonder how hard it is to 'fix' these?

The one about castling is of course almost impossible to cause in a real game, but just in case the programmers ought to be able to turn off tablebase use if castling is still possible. I have no idea how you could solve the fifty-move rule one, though.

Very interesting.