CWE Top Most Dangerous Software Weaknesses(cwe.mitre.org) |
CWE Top Most Dangerous Software Weaknesses(cwe.mitre.org) |
* Out-of-bounds Write
* Improper Neutralization of Input During Web Page Generation ('Cross-site Scripting')
* Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection')
* Use After Free
You avoid them by using tools that make it difficult or impossible to introduce such vulnerabilities to begin with. Such as modern, memory safe programming languages.
For many decades, carpenters have been educated about table saw safety. But what finally stopped thousands of fingers getting chopped off every year was the introduction of the SawStop, and similar technologies.
Safety is a matter of using the right tools, not of "taking better care".
The reason XSS (and CORS) are tricky is because they fundamentally don't work in a world where a website may be spread over a couple different domains. I get a taste of this in my dayjob where we have to manage cookie scoping across a couple different region domains and have several different subdomains for different cookie behaviors. It's easy to be clean on paper up until you need to interface with some piece of software that insists on doing it its own way - for example the Azure excel embedded functionality requires the ID token to be passed in the request body, meaning you have to pull in the request body and parse it in your gateway layer (or delegate that to a microservice)... potentially with multi-GB files being sent in the body as well!
It's super easy on paper to start from greenfield and design something that is sane and clean, bing boom so simple. But once you acquire a couple of these fixed requirements, the cleanliness of the system degrades quite a bit, because that domain uses a format that's not shared by anything else in the system, and it's a bad one, and we can't do anything about it, and now that's a whole separate identity token that has to be managed in parallel.
Anyway, you could say that buffer overflow or use-after-free are kind of an impedence mismatch for memory management/ownership in C. Well, XSS and CORS are an impedence mismatch for domain-based scoping models in a REST-based world. Obviously the correct answer is to simply not write vulnerable systems, but is domain-based scoping making that easier or harder?
Actually the SQL one is arguably in that category too, to a lesser extent. Libraries could, and should, make it obvious how to do parametrized SQL queries in your language. I would guess that for every extra minute of their day a programmer in your language must spend to get the parametrized version to work over just lazy string mangling, you're significantly adding to the resulting vulnerability count because some of them won't bother.
Bonus points if your example code, which people will copy-paste, just uses a fixed query string because it was only an example and surely they'll change that.
Kids get into it just by having the tenacity to do whatever it takes to make it chooch. It's all that counts.
This is often ignored as it simply takes too much time and it often does not hurt much as it’s ‘internal’ (to the company using the saas or whatever).
https://security.googleblog.com/2021/02/mitigating-memory-sa...
https://www.chainguard.dev/unchained/building-the-first-memo...
(Not in use rust for everything bandwagon, genuinely curious)
Either way, they're effectively "solved" from a programmer's perspective if you're willing to adopt modern frameworks instead of string-concatenating HTML or SQL manually.
The middle two are out of reach of a typical PL or type system (there are exceptions like Ur, but I don't think it's adopted widely). It's a problem that is typically solved via libraries and Rust is not unique in terms of providing safe libraries around generating SQL or HTML.
for 4 the static analyzer should help, and, also set your pointer to NULL immediately after free too(for double free)
Changing everything to take lengths is definitely a good change - but challenging to retrofit into existing codebases. Apple has a neat idea for automatically passing lengths along via compilation changes rather than source changes, but if you want to do things in source you have to deal with the fact that there is some function somewhere that takes a void*, increments it locally, reinterpret_casts it to some type, and then accesses one of its fields and you've got a fucking mess of a refactor on your hands.
Use after free is actually gaining popularity, up 3 since last year.
- Improper Neutralization of Input During Web Page Generation ('Cross-site Scripting') (#2)
- Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection') (#3)
- Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection') (#4)
- Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') (#8)
- Improper Neutralization of Special Elements used in a Command ('Command Injection') (#16)
- Improper Control of Generation of Code ('Code Injection') (#23)
All of these came from trying to avoid structured data, and instead using strings with "special characters". It's crazy how many times this mistake has been repeated: file paths, URLs, log files, CSV, HTML, HTTP (cookies, headers, query strings), domain names, SQL, shell commands, shell pipelines... One unescaped character, from anywhere in the stack, and it all blows up.
One could say "at least it's human-readable", but that's not reliable either. Take files names, for example. Two visually identical file names may map to different files (because confusables[1] or surrounding spaces), or two different names map to the same file (because normalization[2]), or the ".jpg" at the end may not actually be the extension (because right-to-left override[3]).
So the computer interpretation of a string might be wrong because a special character sneaked in. And even if everyone was perfectly careful, the human interpretation might still be wrong. For the sake of the next generations, I hope we leave strings for human text and nothing more.
[1] https://unicode.org/cldr/utility/confusables.jsp
[2] https://developer.apple.com/library/archive/qa/qa1173/_index...
[3] https://krebsonsecurity.com/2011/09/right-to-left-override-a...
Thank god Spring dropped this interface in the Framework 6.x / Boot 3.x release, and the end for non-commercial support is this year for the old stuff.
https://github.com/spring-projects/spring-framework/issues/2... https://github.com/advisories/GHSA-4wrc-f8pq-fpqp
RCE via deserilaization seems valid 9.8 even if it requires the developer to use less common APIs or using them in strange ways. In the bug they have a comment that the documentation warns about these API but that doesn't really impact a CVSS score. Am I missing something about this specific CVE on why you think its unfair?
1. CWE-787 Out-of-bounds Write: C, C++, Assembly
4. CWE-416 Use After Free: C, C++
7. CWE-125 Out-of-bounds Read: C, C++
10. CWE-434 Unrestricted Upload of File with Dangerous Type: ASP.NET, PHP, Class: Not Language-Specific
12. CWE-476 NULL Pointer Dereference: C, C++, Java, C#, Go
15. CWE-502 Deserialization of Untrusted Data: Java, Ruby, PHP, Python, JavaScript
17. CWE-119 Improper Restriction of Operations within the Bounds of a Memory Buffer: C, C++, Assembly
21. CWE-362 Concurrent Execution using Shared Resource with Improper Synchronization ('Race Condition'): C, C++, Java
23. CWE-94 Improper Control of Generation of Code ('Code Injection'): Interpreted
I've noticed that using Valgrind on Python systems is almost impossible because most modules have not been built with Valgrind in mind and thus you get swamped in noise.
I suppose the same is true for any large system that uses many different third party libraries.
I'm also not sure if asan has an equivalent to --leak-check=full
I have even made it recognize my custom allocators and report bugs with them too.
When combined with my second favorite tool, AFL++, I have a good shot at eliminating most memory bugs. AFL++ finds paths through the software, and I run every single one of those paths through Valgrind. It's beautiful.
From the optimistic side, it looks like the safest language to write an app today with is TypeScript.
But yes, URLs should have been structured. We already see paths rendered with breadcrumbs, the protocol replaced with an icon, `www` auto-inserted and hidden, and the domain highlighted. If that's not a structure, I don't know what is.
By cramming everything into the same string, we open ourselves to phishing attacks by domains like `www.google.com.evil.com`, malicious traversal, 404s from mangled relative paths, and much more.
Two visually identical file names may map to different files (because confusables[1]), or two different names map to the same file (because normalization[2]), or the ".jpg" at the end may not actually be the extension (because right-to-left override[3]),
Those are all because of Unicode, which is an even worse idea in general.
> it's just that for some reason (like basic arithmetic) people don't seem to be taught enough about it to understand.
That's the same argument used to defend manual memory management. But education is not enough. Escaping is something you have to remember to do every time*, or it'll blow up spectacularly. Even knowledgeable professionals mess it up, or it wouldn't occupy 6 of the 25 spots in this list.
> Those are all because of Unicode, which is an even worse idea in general.
What's the alternative? Japanese speakers writing file names in ASCII? Unicode is a modern marvel, it's our fault we use it where it doesn't belong.
* Not necessarily every input/output, but at least every system that interacts with it.
We make it look like it is a request response with a chat bot, but it is more realistic to say we are making a single document and having the model fill out the rest. That is, there is no out of band. There is only the document.
In java you'll get an exception, while in C you might dissapear your cat. Those 2 are quite incomparable when talking about "dangerous-ness" of a mistake
> 21. CWE-362 Concurrent Execution using Shared Resource with Improper Synchronization ('Race Condition'): C, C++, Java
Why those languages specifically? I would say these two issues apply to all languages.
I understand memory safety is important, but still: only one in the podium (though it is first), only 3 in the top 10… clearly security is about much more than memory safety.
But it is embarrassing that we've been living with memory safety issues for 50 years and they still remain very common and very severe, despite being addressable via type systems in ways that something like a logical bug that leads to data leakage isn't.
To the extent that memory safety is slowly, oh so slowly, but steadily dropping down the list, it is because we are taking it seriously as a foundational issue and actually addressing it. To turn around and then use the success we've had as evidence that it isn't important is making a serious error.
There is no reason to use a memory unsafe language anymore, except legacy codebases, and that is also slowly but surely diminishing. I'm still yet to hear this amazingly compelling reason that you just need memory unsafe languages. In terms of cost/benefits analysis, memory unsafety is literally all costs. Even if you do have one of the rare cases when you need it, and you only need a very particular variant of it (reading bytes in memory of one type as bytes of another type, you never need to write out of the scope of an array or dereference into an unallocated memory page), you can still get it through explicit unsafe support that every language has one way or another. You do not need a language that is pervasively unsafe with every line you write so that on those three lines of code out of millions that you actually need it, you can have it with slightly less ceremony. That's just a mind-blowingly bad tradeoff and engineering decision.
How are we supposed to address the other issues from a foundation of a memory unsafe language? If we can't even have such a basic guarantee, we sure aren't going to get more complicated ones later.
If you don't think about the other classes, I'm still gonna escalate privileges, root your box, ransom your data, send spam, charge a half million dollars in cloud spend to your account, steal your customers' PII/PHI, etc etc etc. Without ever using a language specific exploit.
Escaping by default has become a standard practice with HTML templating languages, see the Go html template standard library for a very detailed breakdown of what is escaped where.
More modern PHP frameworks like Laravel provide their own templating solution in part because of this. But the vast majority of websites run on default PHP templates, so it's not surprising that these kinds of vulnerabilities are so high up in the list.
The whole problem is that you mix code and data, and that third party resource loading is 'on' by default in browsers, especially for scripts and things that can embed scripts. This is not something you can fix once and for all at the library level.
I know it's the nuclear option, but decades of experience has shown that the wider industry just cannot be trusted. People won't ever change[1], so the tools must change to account for that.
[1] Unfortunately, LLMs learned from people... so... sigh.
However, existing libraries for statically-typed languages often don't do the work or apply the creativity and end up roughly as unsafe as the dynamically typed languages.
It's a bit of a pet peeve of mine.
Started with the publishing of "Modern C++ Design" from Andrei Alexandrescu in 2001.
https://en.wikipedia.org/wiki/Modern_C%2B%2B_Design
When ISO C++11 came to be, many re-used the term to mean C++11 or higher.
Given that many keep updating this to mean more modern versions, a well known developer in the community (Tony Van Eerd) has made the joke of that by C++17 time we were in Postmodern C++.
https://www.youtube.com/watch?v=QTLn3goa3A8
No idea what kind of modernism to call C++23, when C++17 was already postmodern, maybe Revivalist C++.
However it basically comes back to Andrei Alexandrescu's original ideas of programming in C++ as its own language, leave the C ways and pitfalls of resource management behind, learn to embrace a modern language for systems programming.
I should also note that there are developers against this philosophy, they advocate that the C++ as understood by CFront is what one should care about, thus Orthodox C++ movement was born.
I'm with Kate Gregory on the "Stop teaching C" (actually Kate specifically means in order to then teach C++ but I also think it's probably fine to stop teaching C outright)
But whilst Kate is right in terms of pedagogy, as a larger philosophy this is inadequate. As a language C++ is obviously defective and the explanation is almost invariably "Because C" which only makes sense once you appreciate C++ in terms of C.
The built-in array type in C++ is garbage. Why is it garbage? This is a language with all these powerful features, why doesn't its array type leverage any of them? It's because this is actually the array type from C.
OK, maybe just the array type is trash, that's obviously not good, but it's one defect. How about string literals. Oops. C++ does sort of technically have the string literals you actually wanted, but the syntax for them is weird and you need the standard library not the core language... the ones you get for "Some text" are C's constant strings, an array of bytes with an extra zero byte, and well, the array type sucks.
This carries on, the language doesn't provide real tuples, it doesn't provide a real sum type, its built-in types don't believe in methods but user types do, everywhere there are weird choices which are non-sensical except for the reality that it's what C does.
And then at the end of that, the language isn't actually compatible with C. It's close, a lot of stuff works, and more stuff kinda-sorta works enough that you may be surprised when it fails, but there isn't the sort of robust compatibility you might expect given the enormous sacrifices made for this goal.
The issue is how "worse is better" culture tends to win, and if the option is between C and C++ for a given scenario, then I definitely take C++.
However if the option pool is widened to more alternatives, then yeah, there should be a sound reason for still pick them for greenfield development, e.g. CUDA, a language toolchain based on LLVM,...
It's somewhat similar to the C/C++ split. When it is convenient it's "C/C++" because "you can easily migrate your old C codebase to C++". But in other situations it's "C++", because C is old and more error prone and "we no longer manipulate raw pointers".
I don't see how you can get away from having a defined serialisation format. People try to operate directly on the serialised data using ad-hoc implementations and run into trouble.
But I'm not sure exactly what you mean by "should have been structured". Eventually you've gotta define the bytes if you want to interoperate with other software.
Yep, that's exactly it. Your TLS certificate is not sent as string, and neither are your TCP packets, nor the images contained in them. Your URLs shouldn't be either, but it's probably too late for that.
> People try to operate directly on the serialised data using ad-hoc implementations and run into trouble.
That's a whole lot better than the current footgun we have, where
http://http://http://@http://http://?http://#http://
> Yep, that's exactly it. Your TLS certificate is not sent as string, and neither are your TCP packets, nor the images contained in them.
...all of those things mentioned have defined serialization. i expect all of them have had security issues because of problems with deserialization code.
That is, it is easy to see that the reason we have URLs sent as strings, is that we collect them from the user. And it makes perfect sense that we would collect strings of characters from users.
What is your point? That strings don't need defined formats? That they have less security issues?
Don’t get me wrong, I absolutely hate the insanity of Undefined Behaviour™ in C and C++ (my pet peeve being signed integer overflow), and I’m totally behind systematic bounds checking (which with compile time support tends to lie between free and cheap). I’m less sold on ensuring the safety of memory shared between threads because I tend to prefer message passing, and I’m not sure how to best address use-after-frees: using the general allocator for each and every object is often even more wasteful than just using a GC, so RAII based schemes aren’t quite enough. I have yet to really test Rust’s borrow checker however.
One thing I have noted, is that C, despite its expressive weakness and its unsafe insanity, remains pretty capable at some niches. Low-level cryptographic code for instance is hardly affected by its flaws (having no heap allocation and constant time code helps a ton).
Memory safety, at least to my eyes, has not traditionally encompassed that as a requirement. I don't consider this a solved problem, in that it has a lot of solutions and consensus about them is still developing. (e.g., I still expect async as it has been implemented in Node & Rust to eventually be considered a gigantic mistake but clearly that is not an uncontroversial opinion in 2023; check in with me in 2033 or 2043). So I'd advise trying to use one of the better solutions but I'm not quite to "there's no reason to not use one of these things".
So my passion is mostly about out-of-bounds access and use-after-free. If it costs you performance... take the hit. It's not a lot. And if you do need unsafe approaches, they are almost always some tight loop somewhere or something where you can selectively take the gloves off and drop down to assembler or something. You don't need you entire language to be unsafe just so you don't have to wrap "unsafe { }" around your tight inner loop.
Yeah, those are the big ones indeed, and I am willing to take a performance hit to get there. If that’s the only hit I take I’ll still be much better than paying an Electron tax.
I do however still feel some discomfort about use-after-free, because to be honest I just don’t know enough about the relevant use cases, compilation techniques, and runtime checks. So far my only relevant experiences have been GC, RAII, and stack-only. They all solve my problem (or at least I can see how I could write a compiler that would solve each use case for me). But I know those aren’t the only use cases, and I’m not familiar enough with the other allocation patterns (pool, arena…) to have a relevant opinion.
But perhaps I’m just stressing over nothing? The problem is easily stated after all: no object should be accessed after its backing memory has been freed. One way to do that is to make sure the object (and any reference to it) goes out of scope before the backing storage is freed. Which sounds doable enough if the backing storage itself follows a stack discipline…
Hey, I can glimpse here a way to allow allocations and statically guarantee a limit on memory usage (barring input dependant allocation amounts). Perhaps even avoiding fragmentation, which would be terrific for embedded use cases.
Tell that to the authors of new memory unsafe languages (like Zig) and creators of new projects in those languages (like https://tigerbeetle.com) :(
It is a huge uphill battle to become a new general purpose language, and the smallest thing can kill it. The fact that Zig is memory unsafe means that I, who am not a bleeding-edge adopter, but am an early adopter and in a position to make decisions about what is used at work, have disqualified it and lost all interest. I have no use for such a language for greenfield projects. Simply offering a more convenient onramp to the sorts of problems that C has is not a compelling value proposition for me.
I extremely strongly suspect I am not even remotely alone.
And in those rare cases, of course you can collect strings from the user. But then they have to be parsed, and that's what should be on the wire. IP addresses are also sometimes entered by hand, but we don't send those strings in TCP packets.
Coverity exists. They've got a great product. But it doesn't solve the problem.
An argument for more often: C++ is so complicated, maybe you're too busy with other problems to address the neutralization issue
An argument for less often: C++ teaches you to be careful and check everything to avoid nasty outcomes so that carries over to neutralization
Afaik the technology isn’t widespread and there are still 10s of thousands of injuries per year.
I will take this bet.
No. All tools can be used with an improper attitude which leads to the creation of weak points.
The proper way is to have a deep understanding of the role of design rules.
A programmer who does not pay attention to design (very basic principles of the design process) can create a good game, and even if this game contains weaknesses the risk related isn't a reason to not use it. The same programmer when creating critical infrastructure software is a source of potential nightmare.
Unfortunately, software business accepts such specialists for projects both of kinds. Why? Who knows? Perhaps because of legal regulations? Why when an engineer designs a car they don't try to "Move fast and break things"?
Similarly, I imagine that there are cases where someone needs to do complex wood working tasks that involve dangers which are a less obvious than with a table saw.
I love static analysis. I did my PhD in it. But we'll still be talking about use after free in 2073 if we just try to chase higher K in our analysis implementations.
The main issue is the community sub-culture of not adopting tooling as it isn't perfect 100% of the time.
Many of the C++ security conscious folks end up being polyglot, as this subculture eventually wears one out.
Then there is the whole issue of making it more interesting to look elsewhere instead.
When a door is locked I can still break in by throwing a rock to the window, yet most people do lock the door nonetheless, while most thieves only bother to break the window if there is anything actually valuable in doing so.
"Although we entertained occasional thoughts about implementing one of the major languages of the time like Fortran, PL/I, or Algol 68, such a project seemed hopelessly large for our resources: much simpler and smaller tools were called for. All these languages influenced our work, but it was more fun to do things on our own."
-- https://www.bell-labs.com/usr/dmr/www/chist.html
Then source tapes with an almost symbolic license price for its time, and a commentary book did the rest.
They do when they design submersible or rockets
Again, even if not perfect, and doesn't cover all use cases, what about people would actually use something at all?
During that decade, how much time did you spent looking at the human side of the problem instead of what the tools can achive?
My only point is that without a different approach we will continue to have the same problems in 2073.