Security breaches in software are really just a special case of software failure, but it has the dubious honour of having a "community" build around it. There is a cost associated with all software failure, but most people are willing to tolerate a certain amount of failure in their applications - as I'm often reminded whenever I discuss formal methods with other programmers - "mostly working is good enough". I can understand this viewpoint (although I don't agree with it), however by this reasoning, what is the acceptable level of security vulnerability in software?
You can tolerate your word processor crashing every few days, but is it acceptable for your entire customer database to be stolen once every few days? How about your production server being rooted once a month? A crippling worm attack once per year? Where does the cost/benefit relationship come into play here?
I don't believe that anyone would seriously advocate the position that a serious security breach once per year resulting in data loss or theft is "acceptable". I doubt anyone wants to have to pay for system administrators to reinstall compromised machines once a month.
The title of this article is "insecure software is your fault". I use "your" to refer to developers. I'll expand on this towards the end of this article, but basically, I believe this to be the case because we have the tools to eliminate most kinds of software vulnerabilities, and yet they still exist, because programmers continue to not employ any of these tools or techniques. If you aren't a software developer, or you actually use the best tools and practices available to you to produce good software, then you can safely assume that I don't mean that insecure software is your fault.
I should create some distinction here between two different types of software "insecurity":
1) Flaws and defects caused by language-level features such as manual memory management or insufficient sanity checks on input that permit damage to data or unauthorised access at the same privilege level as the program itself. SQL injection and some buffer overflow attacks would seem to be obvious examples of this kind of thing.
2) Flaws or defects caused by deficiencies in design - i.e., the software is unable to be compromised itself, however it enables undesirable actions within the application itself. Cross-site scripting attacks are the most obvious example of this - the browser itself remains uncompromised, but any data inside the browser is basically at risk due to an inadequately designed security model.
In the first class of vulnerabilities, there are language-based solutions. Type safety guarantees are sufficiently strong to prevent most kinds of attacks that rely on runtime type violations to exploit flaws. Removing manual memory management from languages reduces another huge class of possible errors. Automatic bounds-checking is the third and final part of it. Contrary to popular myth, the performance cost of these things is very small, and yet the benefit is huge. You've just eliminated even the potential to inadvertently expose yourself and your users to a huge class of security vulnerabilities that will cost time and money to fix.
Claiming that you don't need any of these catastrophe-preventing measures because you're careful is a weak excuse. It's a weak excuse because every experience we've had shows us that being careful just isn't good enough. Suggesting that any developer who has ever introduced a security flaw just wasn't careful enough is a hugely arrogant assertion. If you accept that security vulnerabilities can happen to people who are both smart and careful, then why can't they happen to you? You have a responsibility to your users to be more than careful, you have a responsibility to be sure. And using language-based features that remove a huge class of potential errors is a way to be sure.
In some rare circumstances where performance concerns actually dictate that you must not use these features, don't write a network-facing service. The only place I suggest being careful is in isolating potentially insecure software from the outside world.
There is a small caveat to this - not all languages were created equal. In some cases the design of the language may be sufficient to remove a large class of vulnerabilities (through automatic memory management, bounds checking and sensible typing strategies), however in the interpreted case, you are essentially at the mercy of the interpreter. Your code may be perfectly sane, and the language may be sane, but the interpreter written in C isn't necessarily sane or safe. I add this only as a caution, because the same could be said of the operating system or even the hardware platform - you're only as secure as the things you rely upon. However, anecdotally at least, it's very rare that operating systems have vulnerabilities that are likely to be exploitable only within your application if your application is otherwise defect-free, whereas this isn't always the case with network-facing interpreters.
The second class of errors I identified is errors caused by flaws in design. These can involve things like leaving scope for protocol violations, insufficient security models and some downright stupidity. Very few of these situations have an obvious "language-based" solution at this moment. If your application-level security model is flawed, the language-based features aren't going to help. If illegitimate users of the application are able to cause mischief without ever "exploiting", then the security model is flawed. This could be as simple as leaving a "delete" action outside of some authenticated realm (allowing unauthenticated users to delete things they shouldn't be able to), or it could be that a specially crafted series of actions leads to an undesirable or inconsistent state.
Whatever the specifics, these are all essentially safety properties, as in "this software is not able to exhibit these undesirable behaviours". We actually have ways of checking safety properties nearly automatically these days. You can check various properties through model checking. You can build things inside theorem provers and prove safety properties. Tools like Daikon can even aid in the generation of various invariants through runtime analysis. ESC/Java2 can check certain kinds of safety properties at compile time. BLAST works on C programs. Spec# can express various safety properties, and there are hundreds more similar efforts.
Basically what I'm getting at is that these tools really aren't impossible to use, and they can give you concrete guarantees where you might otherwise have just been guessing. The main thing is the ability to ensure a globally consistent view of the implications of any design decision relating to security. I've previously asserted that humans aren't very good at keeping globally consistent mental models of complex systems, hence why I (and many others before me) advocate mechanical approaches to ensuring global consistency.
So the take-home point from this (rather long) article is that we actually have the tools right now to eliminate most of the kinds of security vulnerabilities that exist. The problem is that people aren't really using them in many places where they really should be. The much-reported damage caused by viruses and worms and "crackers" and DDoS attacks and zombies are all the direct result of insecure software. Sure, people who engage in those sorts of activities are criminals, but really, we're basically leaving our doors and windows wide open and yet we still seems surprised when somebody comes in and steals stuff. If anything, I think we should all start to mentally substitute the phrase "damage caused by cyber-criminals" with "damage caused by irresponsible software developers".
If you happen to be a "computer security professional" (scare quotes intentional), then perhaps you can educate me as to why anti-virus software and firewalls and chasing endless exploits through ad-hoc "penetration testing" is still regarded as the primary activity of the "security community", rather than actually making serious efforts towards preventing security defects in the first place. I have my own pet theory (spoiler alert: $$$), but I'd be interested to hear more informed opinions.
As a side note, I can now announce that I've accepted a PhD position in the PLS group at the IT University of Copenhagen. I'm currently looking for a place for my parter and I to live in Copenhagen from the start of November, so if anyone from Denmark reading this has any leads, please let me know!