12 ms·
What bothers me about this discussion is the assumption that crappy software people are used to is the only type it's possible to build. A capability system wo
by stcredzero 17y ago
What bothers me about this discussion is the assumption that crappy software people are used to is the only type it's possible to build.
A capability system would be a lot harder to completely corrupt. Sandboxing can limit damage to manageable compartments, and enable the system to roll back without user intervention. I think you could make an OS two orders of magnitude harder than current commonly used systems. It will never be complete mathematically provable security. Just make it so hard to do, that the cost increases enough to change the fundamental economics, so that only very high value targets will ever get that degree of attention. (Which means that the cost has to be much higher than someone's credit card number.)
- tptacek 17y agoBernstein disagrees with you, in his analyis of 10 years of qmail security. Worth reading. The problem with sandboxing is the misalignment of effort between attacker and defender. Sandboxing and capabilities are a huge pain in the ass for the defender, who have to construct and deploy applications with perfectly configured security controls. They're just an obstacle course for attackers.
- Retric 17y agoqmail is a program written in C which runs on UNIX. Trying to secure that is a near impossible task. Building a secure system would basically need to start from scratch with a micro kernel which supported a sandboxed OS written in an interpreted language. Even then drivers and HW would need to be treated with the utmost care. Granted, actually building such a system would be horribly expensive and probably take the better part of a decade, but it's still possible.
- tptacek 17y agoIt's easy to specify secure systems that don't do anything in the real world.
- Retric 17y agoI completely agree.
- stcredzero 17y agoneed to start from scratch with a micro kernel which supported a sandboxed OS written in an interpreted language Didn't Microsoft already do this with Singularity? Granted, actually building such a system would be horribly expensive and probably take the better part of a decade, but it's still possible I don't think it would be so horribly expensive. The Lisp machine software was written by a team of 8 or so. The effort for an OS is probably commensurate with BeOS. That's expensive, but not horribly so. Retarget a C compiler for your VM bytecodes, or better yet, just have everything run in an NaCl sandbox. Instead of bytecodes, have your virtual ISA be pseudo x86 or ARM instructions. Then port a lot of GNU code over.
- tsally 17y agoqmail is a program written in C which runs on UNIX. Trying to secure that is a near impossible task. What are you talking about? There hasn't been a published security flaw in qmail yet.