Ideally we'd just keep each mod in a separate lua_State and make mods pass data amongst themselves through a Minetest API. This would improve security and probably make mod-level multithreading easier, but it would break just about all mods. Without this we're limited to a "best guess" as to what mod is running. During init time this guess is fairly accurate, but after init time it's a mess.
For example, even at init time I think the following sandbox escape is possible (although I haven't actually tested it):
minetest.conf:
secure.trusted_mods = trusted
mods/trusted/depends.txt
untrusted
mods/trusted/init.lua:
untrusted()
mods/untrusted/init.lua:
function untrusted() assert(minetest.get_insecure_environment()) --[[ !!! ]] end
The less extreme case of an untrusted mod writing to another untrusted mod's files is also possible (but less problematic since there's no outright sandbox escape).
This issue isn't possible if your trusted mods are written carefully (or better yet, you have no trusted mods), but it's quite a pitfall.
And mod interactions become much more complicated after init time. Minetest has a "best guess" for debug purposes (see core.{get,set}_last_run_mod()), but it isn't reliable enough to bother checking it in a security context.
In short: adding security over an API that was never designed to be secure is hard, and all of these compatibility hacks are getting annoying. :-P