I disagree about the short term bump part.

Quant trading algorithms, like most stochastic trading approaches, lose their effectiveness over time as the market internalizes them and as knowledge dissipates naturally (not necessarily through leaks). The best of the best individual quant trading algorithms have a half-life that does not exceed one month. And we're talking exceptional here. Also, at any given time many, many algorithms will be running simultaneously; complementing, overriding eachother as necessary.

So there's a lot of churn overall. The code that someone picks up in this way contains a single snapshot of this transient trading strategy, not particularly useful to run on their own. You're right that, what really sucks for Goldman is that someone could find some basic patterns inherent to their trading platform and trade "against" Goldman.

The catch is that this sort of maneuvering happens all the time. Sure it's going to be painful in the near term for the Goldman quant guys, but they don't need to start from scratch necessarily.

I think it can last longer than a couple months. I worked on some relative value trading algorithms in connection with work that a b school prof was working on, which got picked up by some of the major HF players. It continued to work for a couple years before the correlations faded away.

Relative value and high-frequency trading are very different animals. With relative value, you probably don't even know if you have statistically significant results for a couple months; with HF, you have evidence within hours.

Do these things really work, or are they random and software validates a belief in them?

It's the low latency that works more than the actual trading algorithm. Algorithms eventually get diluated into the market as there are so many of them.