I think this is one of those domains where structural uncertainty plays a large part. If you're talking about a Bayesian superintelligence operating at the physical limits of computation... I'd feel rather uneasy making speculations as to what limits it could possibly have. In a Tegmark ensemble universe, you get possibilities like 'hacking out of the matrix' or acausal trade or similarly AGI meta-golden rule cooperative optimization, and that's some seriously powerful stuff.
Giulio Prisco made a blog post giving permission to use the data in his Gmail account to reconstruct an uploaded copy of him.
Ben Goertzel copied the post and gave the same permission on his own blog. I made some substantial changes, such as adding a caveat to exclude the possibility of torture worlds (unlikely I know, but can't hurt), and likewise gave permission in my blog. Anders Sandberg comments on the thing.