http://www.kitguru.net/gaming/security-software/carl/webgl-e...
http://www.contextis.com/resources/blog/webgl-more-webgl-sec...
http://www.pcworld.com/article/227438/dangerous_webgl_flaw_p...
My feeling having written a few GLSL shaders is that the language is already sandboxed by design, making it a safer environment to execute 'arbitrary code' than just about any other viable option. In a shader, I get access only to the input I'm given, and can only affect the color of a pixel, and nothing more.
The 'graphics memory theft' exploit (your 2nd link) cannot have been a primary result of a GLSL exploit, as I understand it. That would have required use of an improperly authorized API call like glGetPixels(), and so arbitrary code execution is not the issue there. WebGL's API design and the browser's sandboxing are accountable -- and anyway, it sounds like that got fixed a long time ago?
The first link says the main risk is crashing, so some sort of DOS attack I suppose. That is true of all computing systems, and also isn't exclusive to arbitrary code execution. I'd speculate it's probably easier to get WebGL to crash using the API than shaders. Either way yes, this is a risk, but not particularly specific to WebGL and not the worst kind of security risk in most environments unless combined with something else. (I grant that crashes in hospital equipment, for example, is a high security risk. Then again, most hospitals don't actually let their sensitive equipment browse the web for a reason.)
What other real risks are there, to the average person surfing the web?
There haven't been any real WebGL exploits I know of. No you can't upload arbitrary shaders either. They're validated and then re-written by the browsers.
Every sandboxed app ever has needed native access of some sort to do something useful.
The best example of this difficulty is probably graphics (WebGL). To provide a compelling user experience, the API must allow apps to upload almost arbitrary code (shaders).