I guess I see 0 as the asymptote. Like if you're already at 99.99% reliability, adding another 9 is quite hard. On the other hand, if you're at 10%, then there are big gains to be had.
> "low-hanging fruit" tends to mean doing unprincipled things that can't be generalized / don't contribute anything in the long term, right?
I think sandboxes generalize much better than application-specific proofs. Once you have a provably correct sandbox (which I think is possible today, if you exclude things like 3D acceleration), you can run whatever you want in it: old software, games, Flash, Windows 98. Application-specific proofs only work for applications written in the approved way.
What would a generic sandbox enforce? That an application never accesses the network? That it never accesses the local filesystem? That it never communicates with another process? Browsers need to do all those things and more. I think you need application-specific knowledge to be able to enforce the restrictions that matter.
To some extent, this is a question of what the requirements are. If a sandbox limits a browser to accessing certain files (a la chroot), is that secure? Or does it need to be more fine grained? This isn't something that can be proven, it's mostly a matter of user interface design.
I think there are good arguments for keeping security requirements relatively simple and coarse (including ease of implementation and making sure users can understand what guarantees are offered).
1. It's likely to have bugs because it's mixed with a constantly changing kernel and can't be proven correct
2. It isn't fine-grained enough
If pledge were completely bulletproof, but still limited in terms of what it could restrict, would that still be worthless? Granularity is an interesting problem but it's more subtle than typical criticisms of ad-hoc mitigations.
I think the idea that you can allow a program to get itself in an attacker-controlled state and that will be ok as long as you blacklist what the program can do is fundamentally unworkable. So yes, I think pledge-like approaches are always going to be worthless in the long term - if the program is under the attacker's control, it will almost certainly have enough access to be able to do damage (with a sufficiently skilled attacker), because every program does something, particularly if we're talking about a large and complicated program like a browser. (You could potentially segregate the browser into multiple processes with distinct responsibilities, but I'm not convinced that helps, because those processes still have to send commands to each other and so an attacker who can subvert one can probably control the others).
If it had the level of granularity to express restrictions like "should be able to write only files that the local user has chosen via the save dialogue" then it might become an effective security layer, but at that point we're not really talking about a sandbox any more.
> "should be able to write only files that the local user has chosen via the save dialogue"
The web platform, incidentally, supports this! You can invoke a "choose file" dialog, and then your web page gets access to the one file the user chose.
In capability-based security circles, this pattern is called a "Powerbox". The pattern works especially well in capability systems, where it can be used to choose more than just files.
FWIW, Apple's Mac sandbox and Sandstorm both have open/save support. It's definitely user-friendly and nice polish, but I don't think it's a make-or-break issue.
For anything halfway complex, you can't reasonably find all the bugs for precisely the same reason you can't reasonably get full path code coverage out of your tests. The possibilities around interactions (this is where bugs mostly live) scale very, very high as more components are involved. A simple example is that ten boolean flags or checkboxes that interact with each other--or, more accurately, possibly interact with each other, if you can't prove they don't--means 2^10 possible states to check. Most things don't just take two values. Add system-level and user-level non-determinism and good luck.
In quality, we handle this via testing by state equivalence, predicting how common a set of interactions will be and prioritizing accordingly, and other shortcut techniques meant to put a finite amount of effort around reducing risk to whatever won't sink your product, but it's almost impossible to reduce to zero. You have to spend a metric crapton of money to do it, which is why A) you mostly only see that kind of goal in stuff like aerospace and B) planes cost an awful lot.
I haven't even gotten to the part where you don't know all the possible defective side effects that could be there and probably wouldn't think to look for everything possible even if you theoretically could. Keep in mind that you're treating the application behavior as inherently untrustworthy, so you can't just assume you know how it'd fail either. For all you know it'll interact with an OS bug in a novel way. There are surprises everywhere.
The sandbox idea is a really good one because it means you can do this once, presumably in a system of controlled complexity architected in a way to be highly testable and highly reviewable. That's your best bet for getting near zero defects: high modularity with strict interfaces and limited interactions all of which are highly deterministic. You definitely won't get that with something as complex as an entire browser platform.
Making the job as small and easy as possible and only doing it once are probably the key here.
Huh? That's not how it works, is it? If we want to minimize X * Y and we currently have X = 50 and Y = 5, it's much more efficient to focus on bringing Y down.
> Once we've picked the low-hanging fruit (including perhaps a provably correct sandbox), then we can start thinking about how to prove the correctness of random applications.
"low-hanging fruit" tends to mean doing unprincipled things that can't be generalized / don't contribute anything in the long term, right? My view is that there's limited value in lowering the rate of security flaws in ways that aren't on the path to bringing it to zero. Getting it to zero will be valuable; halving it isn't really (there's some economic value in reducing the frequency of breaches, but not enough). So I don't think ad-hoc countermeasures are worthwhile.
To the extent to which a sandbox can be written in a principled/provable way it will be valuable. I'm not at all convinced that a general-purpose sandbox is possible, but that's a different question. (The techniques of factoring a program into pieces with the minimal capabilities that they need are valuable, but I think this needs to be done far more holistically than is possible with a sandbox; the security design needs to reflect the program design, because whether particular operations are safe or not is a function of the application-specific context. But this is very much speculation on my part)