Go Proposal: Secret Mode

https://antonz.org/accepted/runtime-secret/

by enz • 8 months ago

236 115 8 months ago

fsmv2• 8 months ago

One thing that makes me unsure about this proposal is the silent downgrading on unsupported platforms. People might think they're safe when they're not.

Go has the best support for cryptography of any language

samdoesnothing• 8 months ago

Does it? I'm not disputing you, I'm curious why you think so.

pants2• 7 months ago

Not OP, but Go has some major advantages in cryptography:

1. Well-supported standard libraries generally written by Google

2. Major projects like Vault and K8s that use those implementations and publish new stuff

3. Primary client language for many blockchains, bringing cryptography contributions from the likes of Ethereum Foundation, Tendermint, Algorand, ZK rollups, etc

fastest963• 7 months ago

I'm not sure there's a realistic alternative. If you need to generate a key then it has to happen somehow on unsupported platforms. You can check Enabled() if you need to know and intend to do something different but I assume most of the time you run the same function either way you'd just prefer to opt into secret mode if it's available.

awithrow• 7 months ago

Why not just panic and make it obvious?

kbolino6• 7 months ago

This is not what secret.Enabled() means. But it probably illustrates that the function needs to be renamed already. Here's what the doc comment says:

  // Enabled reports whether Do appears anywhere on the call stack.

In other words, it is just a way of checking that you are indeed running inside the context of some secret.Do call; it doesn't guarantee that secret.Do is actually offering the protection you may desire.

oncallthrow• 7 months ago

Meh, this is a defence in depth measure anyway

Edit: also, the supported platforms are ARM and x86. If your code isn’t running on one of those platforms, you probably know what you’re doing.

ctoth• 7 months ago

Linux

Windows and MacOS?

Go is supposed to be cross-platform. I guess it's cross-platform until it isn't, and will silently change the semantics of security-critical operations (yes, every library builder will definitely remember to check if it's enabled.)

hypeatei• 7 months ago

Meh, this is a defence in depth measure

Which is exactly why it should fail explicitly on unsupported platforms unless the developer says otherwise. I'm not sure how Go developers make things obvious, but presumably you have an ugly method or configuration option like:

  dangerousAllowSecretsToLeak()

...for when a developer understands the risk and doesn't want to panic.

alanfranz• 7 months ago

I'd probably want some way to understand whether secret.Do is launched within a secret-supporting environment so that I'm able to show some user warning / force a user confirmation or generate_secrets_on_unsupported_platforms flag.

But, this is probably a net improvement over the current situation, and this is still experimental, so, changes can happen before it gets to GA.

treyd• 7 months ago

Go has the best support for cryptography of any language

This isn't true at all.

Writing cryptography code in Go is incredibly annoying and cumbersome due lack of operator overloading, forcinforcing you to do method calls like `foo.Add(bar.Mul(baz).Mod(modulus)).Mod(modulus)`. These also often end up having to be bignums instead of using generic fixed-size field arithmetic types. Rust has incredibly extensive cryptographic libraries, the low-level taking advantage of this operator overloading so the code ends up being able to following the notation in literature more closely. The elliptic_curve crate in particular is very nice to work with.

pjmlp2• 7 months ago

Nah, Java and .NET are much better, but they aren't fashionable.

tracker15• 7 months ago

I think C# has been doing really well... I've appreciated the efforts to open the platform since Core... Though I do know a few devs that have been at it as long as I have that don't like the faster lifecycle since the move from Framework.

tracker15• 7 months ago

Guessing it's also taking sure to use assembly calls to zero out and clear the memory region as part of the GC... I would guess the clear/gc characteristics are otherwise the same, but having access to RAM in a non-supported platform could, in theory allow for stale reads of raw memory.

This is likely done for platform performance and having a manual version likely hinders the GC in a way that's deemed too impactful. Beyond this, if SysV or others contribute specific patches that aren't brute forced (such as RiscV extensions), I would assume that the go maintainers would accept it..

dpifke• 8 months ago

Related: https://pkg.go.dev/crypto/subtle#WithDataIndependentTiming (added in 1.25)

And an in-progress proposal to make these various "bubble" functions have consistent semantics: https://github.com/golang/go/issues/76477

(As an aside, the linked blog series is great, but if you're interested in new Go features, I've found it really helpful to also subscribe to https://go.dev/issue/33502 to get the weekly proposal updates straight from the source. Reading the debates on some of these proposals provides a huge level of insight into the evolution of Go.)

kmeisthax• 7 months ago

I have to wonder if we need, say, a special "secret data" type (or modifier) that has the semantics of both crypto/subtle and runtime/secret. That is to say, comparison operators are always constant-time, functions holding the data zero it out immediately, GC immediately zeroes and deallocs secret heap allocations, etc.

I mean, if you're worried about ensuring data gets zeroed out, you probably also don't want to leak it via side channels, either.

maxloh• 7 months ago

The new runtime/secret package lets you run a function in secret mode. After the function finishes, it immediately erases (zeroes out) the registers and stack it used.

I don't understand. Why do you need it in a garbage-collected language?

My impression was that you are not able to access any register in these language. It is handled by the compiler instead.

er4hn• 7 months ago

In theory it prevents failures of the allocator that would allow reading uninitialized memory, which isn't really a thing in Go.

In practice it provides a straightforward path to complying with government crypto certification requirements like FIPS 140 that were written with languages in mind where this is an issue.

kittywantsbacon• 7 months ago

This would potentially protect against other process reading memory via some system compromise - they would be able to get new secrets but not old ones.

jerf• 7 months ago

This is about minimizing attack surface. Not only could secrets be leaked by hacking the OS process somehow to perform arbitrary reads on the memory space and send keys somewhere, they could also be leaked with root access to the machine running the process, root access to the virtualization layer, via other things like rowhammering potentially from an untrusted process in an entirely different virtual context running on the same machine, and at the really high end, attacks where the government agents siezing your machine physically freeze your RAM (that is, reduce the physical temperature of your RAM to very low temperatures) when they confiscate your machine and read it out later. (I don't know if that is still possible with modern RAM, but even if it isn't I wouldn't care to bet much on the proposition that they don't have some other way to read RAM contents out if they really, really want to.) This isn't even intended as a complete list of the possibilities, just more than enough to justify the idea that in very high security environments there's a variety of threats that come from leaving things in RAM longer than you absolutely need to. You can't avoid having things in RAM to operate on them but you can ensure they are as transient as possible to minimize the attack window.

If you are concerned about secrets being zeroed out in almost any language, you need some sort of support for it. Non-GC'd languages are prone to optimize away zeroing out of memory before deallocation, because under normal circumstances a write to a value just before deallocation that is never effectfully read can be dropped without visible consequence to the rest of the program. And as compilers get smarter it can be harder to fool them with code, like, simply reading afterwards with no further visible effect might have been enough to fool 20th century compilers but nowadays I wouldn't count on my compiler being that stupid.

There are also plenty of languages where you may want to use values that are immutable within the context of the language, so there isn't even a way to express "let's zero out this RAM".

Basically, if you don't build this in as a language feature, you have a whole lot of pressures constantly pushing you in the other direction, because why wouldn't you want to avoid the cost of zeroing memory if you can? All kinds of reasons to try to avoid that.

kbolino6• 7 months ago

Go has both assembly language and unsafe pointer operations available. While any uses of these more advanced techniques should be vetted before going to production, they are obviously able to break out of any sandboxing that you might otherwise think a garbage collector provides.

And any language which can call C code that is resident in the same virtual memory space can have its own restrictions bypassed by said C code. This even applies to more restrictive runtimes like the JVM or Python.

vlovich1234• 7 months ago

Go is not a memory safe language. Even in memory safe languages, memory safety vulnerabilities can exist. Such vulnerabilities can be used to hijack your process into running untrusted code. Or as others point out sibling processes could attack yours. This underlying principle is defense in depth - you make add another layer of protection that has to be bypassed to achieve an exploit. All the chains combined raise the expense of hacking a system.

tptacek8• 7 months ago

Respectfully, this has become a message board canard. Go is absolutely a memory safe language. The problem is that "memory safe", in its most common usage, is a term of art, meaning "resilient against memory corruption exploits stemming from bounds checking, pointer provenance, uninitialized variables, type confusion and memory lifecycle issues". To say that Go isn't memory safe under that definition is a "big if true" claim, as it implies that many other mainstream languages commonly regarded as memory safe aren't.

Since "safety" is an encompassing term, it's easy to find more rigorous definitions of the term that Go would flunk; for instance, it relies on explicit synchronization for shared memory variables. People aren't wrong for calling out that other languages have stronger correctness stories, especially regarding concurrency. But they are wrong for extending those claims to "Go isn't memory safe".

https://www.memorysafety.org/docs/memory-safety/

hamburglar2• 7 months ago

The Go runtime may not be the only thing reading your process’ memory.

gethly• 7 months ago

Yeah, I can hardly disagree with that sentiment myself.

tracker15• 7 months ago

That's kind of the point of making it (relatively easily) accessibly... in order to ensure that secrets are wiped from memory as quickly as reasonable. This reduces a lot of potential surface area for potential attack.

I used to not really see the need for this level of detail on things... then you see useful (IE in the wild exploits) for even complex factors like CPU branch prediction (for a while now), and the need starts to become much more clear.

voodooEntity5• 7 months ago

Ok, i kinda get the idea, and with some modification it might be quite handy - but i wonder why its deemed like an "unsolvable" issue right now.

It may sound naive, but packages which include data like said session related or any other that should not persist (until the next Global GC) - why don't you just scramble their value before ending your current action?

And dont get me wrong - yes that implies extra computation yada yada - but until a a solution is practical and builtin - i'd just recommend to scramble such variables with new data so no matter how long it will persist, a dump would just return your "random" scramble and nothing actually relevant.

willahmad• 7 months ago

without language level support, it makes code look like a mess.

Imagine, 3 level nesting calls where each calls another 3 methods, we are talking about 28 functions each with couple of variables, of course you can still clean them up, but imagine how clean code will look if you don't have to.

Just like garbage collection, you can free up memory yourself, but someone forgot something and we have either memory leak or security issues.

compsciphd2• 7 months ago

I could imagine code that did something like this for primatives

  secretStash := NewSecretStash()
  pString := secretStash.NewString()
  ....
  ....
  secretStash.Thrash()

yes, you now have to deal in pointers, but that's not too ugly, and everything is stored in secretStash so can iterate over all the types it supports and thrash them to make them unusable, even without the gc running.

HendrikHensen• 7 months ago

With good helpers, it could become something as simple as

    key := make([]byte, 32)
    defer scramble(&key)
    // do all the secret stuff

Unless I don't understand the problem correctly.

skywhopper• 7 months ago

Hard to understand what you’re asking. This is the solution that will practical and built-in. This is a summary of a new feature coming to Go’s runtime in 1.26.

raggi4• 7 months ago

It is fundamentally not possible to be in complete control of where the data you are working with is stored in go. The compiler is free to put things on the heap or on the stack however it wants. Relatedly it may make whatever copies it likes in between actions defined in the memory model which could leak arbitrary temporaries.

to11mtm• 7 months ago

Yeah, .NET tried to provide a specific type related to this concept (SecureString) in the past and AFAIK there were were two main problems that have caused it to fall into disuse;

First one being, it was -very- tricky to use properly for most cases, APIs to the outside world typically would give a byte[] or string or char[] and then you fall into the problem space you mention. That is, if you used a byte[] or char[] array, and GC does a relocation of the data, it may still be present in the old spot.

(Worth noting, the type itself doesn't do that, whatever you pass in gets copied to a non-gc buffer.)

The second issue is that there's not a unified unix memory protection system like in windows; The windows implementation is able to use Crypt32 such that only the current process can read the memory it used for the buffeer.

leoh4• 7 months ago

Kind of stupid it didn’t have something like this to begin with tbh. It really is an incredible oversight when one steps back. I am fully ready to be downvoted to hell for this, but rust ftw.

IshKebab4• 7 months ago

Rust doesn't have anything like this either. I think you misunderstood what it is.

raggi4• 7 months ago

It doesn’t but the problem space is more constrained as you are at least in control of heap vs stack storage. Register clearing is not natively available though. To put it more simply: yes but you can write this in rust- you can’t write it in go today.

leoh4• 7 months ago

Okay, fair point, sort of. Rust does not have a built-in feature to zero data. Rust does automatically drop references to data on the heap. Zeroing data is fairly trivial, whereas in go, the issue is non-trivial (afaiu).

  use std::ptr;
  
  struct SecretData {
      data: Vec<u8>,
  }
  
  impl Drop for SecretData {
      fn drop(&mut self) {
          // Zero out the data
          unsafe {
              ptr::write_bytes(self.data.as_mut_ptr(), 0, self.data.len());
          }
      }
  }

teeray• 7 months ago

I wonder if people will start using this as magic security sauce.

tjpnz• 7 months ago

More likely they'll use it and end up with a false sense of security.

jeffrallen• 7 months ago

Wow, this is so neat. I spent some time thinking about this problem years ago, and never thought of such an elegant solution.

raggi4• 7 months ago

This seems like it might be expensive (though plausibly complete), so I wonder if it’ll actually benchmark with a low enough overhead to be practical. We already struggle with a lack of optimization in some of the named target use cases - that said this also means there’s space to make up.

hamburglar2• 7 months ago

Personally, I’m more interested in what a process can do to protect a small amount of secret material longer-term, such as using wired memory and trust zones. I was hoping this would be an abstraction for that.

burnt-resistor• 7 months ago

Consumer-grade hardware generally lacks real confidentiality assurance features. Such a software feature implemented in user-space is moot without the ability to control context switching, rendering it mostly security theater. Security critical bits should be done in a dedicated crypto processor that has tamper self-zeroing and self-contained RAM, or at the very least, in the kernel away outside the reach of user-space processes. No matter how much marketing or blog hype is offered, it's lipstick on a pig. They've essentially implemented a soft, insecure HSM.

Big thumbs down from me.

Someone• 7 months ago

FTA: “Heap allocations made by the function are erased as soon as the garbage collector decides they are no longer reachable”

I think that means this proposal adds a very specific form of finalisers to go.

How is that implemented efficiently? I can think of doing something akin to NSAutoReleasePool (https://developer.apple.com/documentation/foundation/nsautor...), with all allocations inside a `secret.Do` block going into a separate section of the heap (like a new generation), and, on exit of the block, the runtime doing a GC cycle, collecting and clearing every now inaccessible object in that section of the heap.

It can’t do that, though, because the article also says:

“Heap allocations are only erased if the program drops all references to them, and then the garbage collector notices that those references are gone. The program controls the first part, but the second part depends on when the runtime decides to act”

and I think what I am thinking of will guarantee that the garbage collector will eagerly erase any heap allocations that can be freed.

Also, the requirement “ the program drops all references to them” means this is not a 100% free lunch. You can’t simply wrap your code in a `secret.Do` and expect your code to be free of leaking secrets.

_cenw• 7 months ago

My guess is that it just uses the existing finalizers and ensures the memory is overwritten.

https://pkg.go.dev/runtime#SetFinalizer

tapirl• 7 months ago

SetFinalizer is deprecated by AddCleanup: https://pkg.go.dev/runtime#AddCleanup

AddCleanup might be too heavy, it is cheaper to just set a bit in the header/info zone of memory blocks.

tracker15• 7 months ago

Same here, though I think that the GC will do a full run for all "secure" data that is clear regardless of GC timing.

tracker15• 7 months ago

My guess is that the GC will eagerly wipe+free "secret" memory first and completely on the pass as opposed to deferring to the next GC cycle as is the normal use that may otherwise happen on longer GC cycles to keep overall performance characteristics happy.

This is just my own speculation, I don't know the internals of Go beyond a few articles I've read over the years. Somewhat more familiar with C#, JS and Rust, and even then I'll sometimes confuse certain features between them.

robmccoll• 7 months ago

This is interesting, but how do you bootstrap it? How does this little software enclave get key material in that doesn't transit untrusted memory? From a file? I guess the attacker this is guarding against can read parts of memory remotely but doesn't have RCE. Seems like a better approach would be an explicitly separate allocator and message passing boundaries. Maybe a new way to launch an isolated go routine with limited copying channels.

cyberax5• 7 months ago

How does this little software enclave get key material in that doesn't transit untrusted memory?

Linux has memfd_secret ( https://man7.org/linux/man-pages/man2/memfd_secret.2.html ), that allow you to create a secure memory region that can't be directly mapped into regular RAM.

_1tan• 7 months ago

Seems neat, anything similar in Java?

nixpulvis• 7 months ago

I looked into this a bit for a rust project I'm working on, it's slightly difficult to be confident, when you get all the way down to the CPU.

https://github.com/rust-lang/rust/issues/17046

https://github.com/conradkleinespel/rpassword/issues/100#iss...

cafxx3• 7 months ago

I find this example mildly infuriating/amusing:

    func Encrypt(message []byte) ([]byte, error) {
        var ciphertext []byte
        var encErr error
    
        secret.Do(func() {
            // ...
        })
        
        return ciphertext, encErr
    }

As that suggests that somehow for PFS it is critical that the ephemeral key (not the long-term one) is zeroed out, while the plaintext message - i.e. the thing that in the example we allegedly want secrecy for - is totally fine to be outside of the whole `secret` machinery, and remain in memory potentially "forever".

I get that the example is simplified (because what it should actually do is protect the long-term key, not the ephemeral one)... so, yeah, it's just a bad example.

kbolino6• 7 months ago

PFS is just one of many desirable properties, and getting access to plaintext is just one of many kinds of threat. Getting access to ephemeral keys and other sensitive state can enable session hijacking. It's still not a great example, though, because it doesn't illustrate that threat model either.

__turbobrew__• 7 months ago

Protection does not cover any global variables that f writes

Seems like this should raise a compiler error or panic on runtime.

jabedude• 7 months ago

How much control does a language runtime have over whether the memory controller actually zeros out physical memory? My guess is very little on consumer hardware but happy to be wrong

circuit10• 7 months ago

If it's no longer readable by software that's at least far better than no protection, I imagine

thr0w4w4y1337• 7 months ago

awnumar/memguard[1] exists and does even more

1) allocations via memguard bypass gc entirely

2) they are encrypted at all times when not unsealed

3) pages are mprotected to prevent leakage via swap

4) and so on...

Not as ergonomic as OP's proposal, of course.

[1] https://github.com/awnumar/memguard

baq2• 7 months ago

    // Only the ciphertext leaves this closure.

This ideally should be describable by the type system.

Thorrez2• 7 months ago

If an offset in an array is itself secret (you have a data array and the secret key always starts at data[100]), don't create a pointer to that location (don't create a pointer p to &data[100]). Otherwise, the garbage collector might store this pointer, since it needs to know about all active pointers to do its job. If someone launches an attack to access the GC's memory, your secret offset could be exposed.

That doesn't make sense to me. How can the "offset in an array itself" be "secret" if it's "always" 100? 100 isn't secret.

stingraycharles• 7 months ago

I think it may be about the absolute memory address to the secret being stored, which may itself be exploitable (ie you’re thinking about the offset value, rather than the pointer value). it’s about leaking even indirect information that could be exploited in different ways. From my understanding, this type of cryptography goes to extremely lengths to basically hide everything.

That’s my hunch at least, but I’m not a security expert.

The example could probably have been better phrased.

Thorrez2• 7 months ago

I don't see how a single absolute address could be exploitable based on my understanding of the threat model of this library. The library is in charge of erasing the secrets from memory. Once the secrets have been erased from memory, what would an attacker gain from knowing an absolute address?

The only thing that makes sense to me is a scenario with a lot of addresses. E.g. if there's an array of 256 integers, and those integers themselves aren't secret. Then there's a key composed of 32 of those integers, and the code picks which integers to use for the key by using pointers to them. If an attacker is able to know those 32 pointers, then the attacker can easily know what 32 integers the key is made of, and can thus know the key. Since the secret package doesn't erase pointers, it doesn't protect against this attack. The solution is to use 32 array indexes to choose the 32 integers, not 32 pointers to choose the 32 integers. The array indexes will be erased by the secret package.

yyyk• 7 months ago

Aka .NET SecureString - which is barely used because everything accepts String.

Go Proposal: Secret Mode

Related Stories