The AdaCore book below illustrates nicely the different areas where defects are common and how they systematically eliminate them. The Wirth languages similarly added all kinds of protections in automatically while letting you disable safety where performance or low-level requirements demanded it. Still did interface protections for calls to that code. I think updating something like that with extra protections Ada community, etc have figured out would give lots of safety, performance, & without Ada's steep learning curve. What you think?
http://www.adacore.com/knowledge/technical-papers/safe-secur...
To the first point: Safety is good. Really. For security-critical software, safety is very, very, important. It's still not the only thing that's important. Familiarity to the developer still matters. Tool support still matters.
For example, I've got nearly 30 years of experience with C/C++. I can avoid (most of) the sharp edges. Could I write code that somebody could still pwn? Probably. Could a static analyzer, say, save me? Maybe. But let's say I wrote the same code in Haskell. Could I write at least something that could DOS the user's machine? Easily. Would I have any idea whether I was doing so? No. Could a static analyzer save me?
Second, I remember when Microsoft added the ability for Outlook to execute attachments. That was going to be a gaping security hole, no matter what language it was written in. The security hole was in the spec, not in implementing language. And if I understand derekp's point, you could have a safety hole in the library. For example, Java is safer than C at string handling... unless there was a flaw in Java's String library. Are you sure that there isn't one? How sure are you? Why are you that sure?
For that matter, did Haskell have an SSL package? Was that package anything besides a wrapper around a C library?
The higher level stuff you use, the more layers there are below you, and the greater the attack surface that you can't even see.
Now, you could argue that I'm more likely to write a security bug doing string handling by hand, than there is to be one in Java's String library. And I'd agree that it's probably so. But I need everything I use to be bulletproof, and an attacker only needs to find one flaw. The more I use, the less the odds are in my favor. (The more I write myself, also the less the odds are in my favor.) There is a place where writing it yourself, even in C, is the right answer, even for security critical code. But once you do, audit it heavily, analyze it with as many tools as you can, have whitehats take a serious run at it, and so on.
Fine, fine, it doesn't "refute" it, it "merely" makes it irrelevant to the real world.
I don't think the rest of your message is particularly relevant to mine. Having claimed that there are varying degrees of safety, I'm not sure how an extended discussion of the varying degrees of safety of various technologies counters anything I said.
(Also, as a pragmatic matter, I consider "C" and "C + Valgrind + Coverity + whatever else" to basically be different languages. However, don't talk to me about how wonderful the latter is if you're programming in the former! You have to actually use the tools, not just get a sense of comfort from the fact they exist.)
If yes to 1, 2: Can you seriously defend the claim that C is not well to the "unsafe" end of the spectrum? And that, equivalently, there are a number of safer languages to write in?
If no to 1, 2: How can you actually justify that? Given the same programmer writing in Haskell and Assembler, with the same amount of effort, all code written by that programmer will have the same security level? Really?
It seems to me the only way to carry the argument that "It's not a problem to write security code in C" is to reduce all languages to "safe" and "unsafe", define "safe" somewhat absurdly as "impossible to write insecure code in", then claim that since all languages are in the "unsafe" column, it's just fine to use C.
The mere act of spelling that argument ought to nearly suffice to refute it.