One Laptop Per Child Security Spec Released

juwiley writes "The One Laptop Per Child project has released information about its advanced security platform called Bitfrost. Could children with a $100 laptop end up with a better security infrastructure than executives using $5000 laptops powered by Vista? 'What's deeply troubling — almost unbelievable — about [Unix style permissions] is that they've remained virtually the only real control mechanism that a user has over her personal documents today...In 1971, this might have been acceptable...We have set out to create a system that is both drastically more secure and provides drastically more usable security than any mainstream system currently on the market.'"

Comments

very sceptical

[Tom]

Security is a lot like crypto: Designing your own system is a recipe for desaster. Security is hard, and aside from the conceptual stages, small failures in implementation can destroy the best concept.

So anyone coming up with a "new and improved" security concept is selling an untested solution. Because security is always tested in the field, never (at least never properly) in the lab.

And yes, Unix permissions are primitive. But they work, they are reliable and we know their shortcomings and limitations.

Re:very sceptical

[swillden]

So anyone coming up with a "new and improved" security concept is selling an untested solution.

True, but inapplicable in this case. For two reasons.

1. There are no new concepts in the XO security model.
2. The traditional security model (used by Unix and Windows) cannot work for the OLPC, so something different is required.

How can we have a new security model, but no new security model concepts? What's new is that ideas which have been reserved for high-security systems are being applied to a system that large numbers of people will actually use.

The core ideas are:

• Sandboxing, aka Mandatory Access Controls. Not only have research systems built on this concept existed for years, but we also have a decade of practical experience with Java sandboxes, and several years of extensive experience with MAC on Linux (SELinux). Specialized high-security operating systems have employed MAC for decades.
• Chroot jails. Most sysadmins who are serious about security run all Internet-facing applications in jails, to limit the damage that can be done if the app is exploited. The only difference here is that the concept is being applied to all apps.
• Digital signatures as a way to authorize applications to break out of their constrained (sandboxed and jailed) environments.
• Allowing users to authorize applications to break out of their constrained environments.
• Security by default. The system is secure out of the box.

The only innovation here is in the decision to apply these known security models/tools to all applications on the OLPC. There is some good thought that has gone into determining what kinds of restrictions can be placed on apps, and the bit about constraining the permissions that apps can request during installation (e.g. either network or file access, not both -- without digital signature or explicit user authorization) is clever, but there's nothing fundamentally new.

But the issue is somewhat deeper than that, as well.

It's important to realize that the traditional security model does not work for OLPC machines. Why? Because (1) they're specifically designed as computers whose software is highly mutable and (2) they're specifically designed to live as part of a network. The traditional model works great if you can thoroughly prove the integrity of the software on the system and then lock it down -- but you can't do that on machines that are constantly connected to others and always exchanging bits of code and data.

You can try, of course. And we do. And we've seen just how well it works. Massive botnets of zombies is the result as is high-powered machines dedicating a significant portion of their processing power to defending themselves against malicious code -- and failing.

The traditional model is fundamentally broken in the networked age, and the OLPC machines are not only networked, but designed to facilitate every user becoming an at least minimally-competent programmer and to encourage widespread, free sharing of user-developed code.

New problems require new solutions. In this case, it appears that we already had all of the tools required available, they just weren't widely used.

My prediction: The XO security model will be an outstanding success story. It'll have its problems, and it'll have to be tweaked in various ways, but the basic ideas are so good, and so fundamentally simple, that it will work very well. Application authors will be able to achieve what they want, and security will be generally quite good.

I also think that the OLPC project is one of the most amazing stories in the history of computing. It's giving a bunch of brilliant people the opportunity to completely re-imagine computing, and they're doing it with a laser focus on the needs of the people who use the computers, rather than the needs of those who sell the computers and the software.

The one major difference to MS "trusted" computing

[gd23ka]

--"No lockdown. Though in their default settings, the laptop's security
  systems may impose various prohibitions on the user's actions, there
must exist a way for these security systems to be disabled. When that is
the case, the machine will grant the user complete control."

That is the one of the key differences between Bitfrost and Microsoft
"trusted computing" schemes: you as owner of the box can get around it.

Re:chmod, chown, etc.?

[pla]

I wonder if the author's used chmod, chown, etc.? What's the essential difference between Unix style permissions and other permission systems?

Well, Windows uses the ACL system of permissions it stole from VMS. It actually does provide more control (that you don't need 99.9% of the time), such as multiple groups having different levels of permissions.

Increasingly complex file-level security does come with one major drawback, however... I can look at a file under Linux and instantly tell (possibly with a quick check of the members of a single group) who has what access to it. Under Windows, good luck with that. XP actually has an advanced security tab, "Effective Permissions", solely for the purpose of testing what access a given user has to a file or directory. Short of that tool, some of the more complex possible configurations (which don't take any sort of unrealistically contrived setups to get, such as a combination of local and domain groups having both inherited and locally set permissions) would leave you feeling very uncomfortable guessing who has access to a given file. And of course, that tab only lets you check one user or group at a time, so it proves utterly useless in answering the simple question "Who can overwrite this file".

In fairness, you could write a script to test every user and group against a given set of files and directories and generate a report off the output, but seriously, would anyone really consider that "better" than "0750, yup, that looks good"?

It's not hard to do this. Just not compatible.

[Animats]

It's not hard to do this. Several groups had systems this tight working back in the 1980s. For that matter, Multics had it right in the late 1960s. Linux has it now, in NSA SELinux.

It breaks existing applications, of course. The OLPC people have a huge advantage - they don't care about existing applications. They can say to application developers, "these are the security constraints - design to them." That's a huge win.

Somebody should have done this by now for phones and palmtops, but, unfortunately, those things started out so underpowered they barely had an operating system. So they have their own legacy problems.

Re:One Desktop per Village would be a better start

[Goaway]

I can't help but notice that the people working on this "too ambitious" project are actually out there doing it, while you are... posting on Slashdot?

It's worse than that, it prevents app partitioning

[Morgaine]

>> how am I going to implement this new idea I have for cross-application communication based on shared pipes among apps.

Actually, it's even worse than your funny (but accurate) comment suggests:

In the Unix model, applications are often built out of multiple cooperating processes, each of which is isolated into its own address space, with strong barriers between processes enforced by the MMU hardware. This makes each separate part more robust, more comprehensible, and more secure.

In contrast, when Bitfrost throws away the ability of programs to talk to other programs, it is intrinsically encouraging a monolithic approach to program design, which is a huge step backwards both for security and for complexity management.

Bitfrost is right to deny free access by programs to a user's filestore objects as an important part of its new security framework, but if the price for that is to disallow strong application factoring and partitioning into separate but communicating processes then the cure may be worse than the disease.

Linux did clone the Plan 9 feature though

[r00t]

Our rfork() is called clone(), or unshare() if you don't need a new thread/process.

When you want a new namespace, you specify the CLONE_NEWNS flag. (root only, sorry, because of setuid concerns)

Once you have a new namespace, you can unmount things you don't need. You can do bind mounts, which let you graft directories onto other places. You can use a bind mount to make a read-only copy of something, then unmount the original... all without mucking up processes that aren't part of the same CLONE_NEWNS group. Portions of the filesystem tree can be shared as well, in case you really do want changes to appear to both sides of the CLONE_NEWNS. Access to things can be permanently given up within the CLONE_NEWNS group, making for a rather fine jail that generally beats jail(8) quite severely.

There are extra goodies for stuff like isolating the view of system time, the view of executing processes, etc.