Using Rust Macros to exfiltrate secrets(github.com)
github.com
Using Rust Macros to exfiltrate secrets
https://github.com/lucky/bad_actor_poc
75 comments
More specifically, a proc macro is a Rust function that is compiled and run inside the compiler at build time. With IDEs, LSP and other protocols for having your editor query the compiler (or language runtime, like SLIME/SWANK), the compiler now runs whenever you open your editor.
It’s just not a new problem. Bash does auto–completion on Makefiles, which requires running make and asking it what the make targets are. IDEs can and will run ./configure for you, so that it can find the right include paths. Etc, etc.
Personally, I thought everyone already knew about this. I knew that proc macros would be a risk when I first heard about rls, years ago.
Certainly editors need to confirm with the user that they are ok with starting the compiler when they load a new project, but also we need to use fine–grained security systems like SELinux that can and do prevent programs from accessing things that they’re not supposed to access.
It’s just not a new problem. Bash does auto–completion on Makefiles, which requires running make and asking it what the make targets are. IDEs can and will run ./configure for you, so that it can find the right include paths. Etc, etc.
Personally, I thought everyone already knew about this. I knew that proc macros would be a risk when I first heard about rls, years ago.
Certainly editors need to confirm with the user that they are ok with starting the compiler when they load a new project, but also we need to use fine–grained security systems like SELinux that can and do prevent programs from accessing things that they’re not supposed to access.
One potential solution:
- During a session, the first time rust-toolchain encounters a proc macro it must run to analyze, it will first prompt the user and warn them.
- If the user accepts the prompt, rust-toolchain will freely run any proc macros until the next session.
- If the user rejects the prompt, that analysis will be disabled until the next session.
Similar to how VSCode and other apps handle opening links.
- During a session, the first time rust-toolchain encounters a proc macro it must run to analyze, it will first prompt the user and warn them.
- If the user accepts the prompt, rust-toolchain will freely run any proc macros until the next session.
- If the user rejects the prompt, that analysis will be disabled until the next session.
Similar to how VSCode and other apps handle opening links.
You'd have to sandbox the analyzer. Let it run arbitrary code but don't let it do IO. That can be pretty tricky to do for a language not designed to be sandboxed.
Safest way would probably be something hilarious like having the analyzer compiled to WASM and ran in node.js.
Safest way would probably be something hilarious like having the analyzer compiled to WASM and ran in node.js.
By default rust-analyzer also executes Rust build scripts (build.rs) just by opening the project in an IDE, so as far as Rust goes the comparison is apt.
rust-analyzer.cargo.runBuildScripts (default: true)
Run build scripts (build.rs) for more precise code analysis.
Open it in notepad? You don't install software that automatically build your project and complain that it is doing that.
Agreed. The top comments on this thread are wrong, overconfident and silly. Read the article people.
This is a huge deal right? VSCode has to be one of the most popular editors and the standard way of setting up the Rust toolchain on a machine would get you in a state that makes you vulnerable to this.
This is as huge a deal as "using ./configure && make install to exfiltrate secrets."
It's a class of supply chain attack focusing on build time code evaluation. Almost every programming language has some kind of support for arbitrary code execution at build time, and any project of scale is going to require it.
RCE isn't an interesting exploit when the system is literally designed to run code from somewhere else.
It's a class of supply chain attack focusing on build time code evaluation. Almost every programming language has some kind of support for arbitrary code execution at build time, and any project of scale is going to require it.
RCE isn't an interesting exploit when the system is literally designed to run code from somewhere else.
This isn’t build time though really, which I agree is a moment you would expect to run arbitrary code. This is “edit time.”
It is build time. Whether rust-analyzer should run build-time code at initialization is a different discussion.
That is a more philosophical definition of build time than what I am referring to.
It's not philosophical, it is literal.
Rust macros and build.rs require build commands to be executed by the rust compiler (cargo check, for example).
These are third party tools that have been implemented to execute build commands during initialization. It's not an issue with Rust, it's an issue with the implementation of the language client and text editor allowing the client to initialize when opening a workspace.
Rust macros and build.rs require build commands to be executed by the rust compiler (cargo check, for example).
These are third party tools that have been implemented to execute build commands during initialization. It's not an issue with Rust, it's an issue with the implementation of the language client and text editor allowing the client to initialize when opening a workspace.
What I mean is that my definition of build time is more things that happen when you are running a build. If you do build time things at other times then I don’t personally consider them build time activities any more. You have converted your build time activities into something more general at that point.
They may be spiritually or metaphorically still build steps in some sense, but they are happening outside that context.
They may be spiritually or metaphorically still build steps in some sense, but they are happening outside that context.
No, it's worse. People will have very different expectations. Running 'make install' especially as root implies a high level of trust, so users will be appropriately cautious. Users won't expect that simply opening code in an editor will be similarly risky (though it's similar to malicious Word and Excel macros, Office now disables those by default for documents coming from an untrusted source, like via email).
The only component is an editor that runs some code automatically.
A python plugin for an editor would have the same problem - if it imports a python module for any reason, like code completion. Same problem of arbitrary code execution.
I think we should work on solutions. Sandboxing both for editor plugins and for regular rust builds, should become the norm.
A python plugin for an editor would have the same problem - if it imports a python module for any reason, like code completion. Same problem of arbitrary code execution.
I think we should work on solutions. Sandboxing both for editor plugins and for regular rust builds, should become the norm.
Sandboxing isn’t necessary. Proper use of SELinux, which the Linux kernel in your computer already supports even if your distro doesn’t enable it, would already prevent any process other than ssh from reading your private key. The build system could run ssh and ssh would be allowed to read the key, but the key is still safe as long as ssh cannot be tricked into revealing it. Since that’s generally believed to be the case, no sandboxing is necessary.
If your distro doesn’t enable SELinux, or your distro’s SELinux policy doesn’t protect your ssh keys, then you need to upgrade. If you don’t use Linux, then you need to upgrade to Linux.
If your distro doesn’t enable SELinux, or your distro’s SELinux policy doesn’t protect your ssh keys, then you need to upgrade. If you don’t use Linux, then you need to upgrade to Linux.
SELinux is nothing but a maintenance burden for developers and users. There's a reason no major distro beyond the Fedora line that enables it by default, including the one I work one; it almost invariably frustrates users, has incomplete support for the workflows they use, and so they turn it off anyway. I literally turned off SELinux on my new ARM64 Fedora machine last week, because it prevented me from installing a third party binary (which I am a developer of.) That binary in turn needed its own ability to use namespacing support to sandbox applications (in a manner that works and offers enhanced security on any Linux distro, not just SELinux ones), etc. It's a non-starter.
> Since that’s generally believed to be the case, no sandboxing is necessary.
That's where you're wrong. It's necessary even if you believe it's not. It's been proven time and time again that this is the case and that the "belief" no flaws exist is wrong.
Sandboxing approaches that use techniques like namespaces, and capability security have become vastly, vastly more popular over the years on Linux, and they're going to keep getting more popular, precisely because they work where SELinux fails (that is, 98% of the running Linux systems and distros that actually exist). Browsers, WebAssembly, systems like Flatpak with "Portals" -- all of them have moved into capability-inspired and "component" sandboxing approaches, to achieve this level of security independent of the host operating system. If Chrome had decided to use SELinux instead of its own sandboxing approach, it's security model would be completely inferior to what it is today.
> Since that’s generally believed to be the case, no sandboxing is necessary.
That's where you're wrong. It's necessary even if you believe it's not. It's been proven time and time again that this is the case and that the "belief" no flaws exist is wrong.
Sandboxing approaches that use techniques like namespaces, and capability security have become vastly, vastly more popular over the years on Linux, and they're going to keep getting more popular, precisely because they work where SELinux fails (that is, 98% of the running Linux systems and distros that actually exist). Browsers, WebAssembly, systems like Flatpak with "Portals" -- all of them have moved into capability-inspired and "component" sandboxing approaches, to achieve this level of security independent of the host operating system. If Chrome had decided to use SELinux instead of its own sandboxing approach, it's security model would be completely inferior to what it is today.
Don’t confuse the sandboxing a browser does with other types. I would never argue that a browser shouldn’t sandbox javascript. The kernel should also prevent the browser from reading your SSH keys as well, just in case :)
If I rememver correctly, the python plugin for vscode asks me whether I trust the project before running anything. At least that was the case when I last opened a Jupyter notebook in vscode.
It requires never actually running the program to be worse than the status quo in any language. If you run code it's trivially a code execution.
This issue is very similar to the problem of malicious macros in Microsoft Office documents, and I think it needs to be addressed somehow (by figuring out a proper security model and asking for user confirmation for actions outside this model).
A lot of the work I do recently has been using devcontainers in VSCode [1]. They even have a Rust sample one. I feel like this would provide at least a little bit of protection against this kind of attack if you do not mount any imporant stuff into the container.
I can't see a robust solution to this, though.
[1] https://code.visualstudio.com/docs/remote/containers
I can't see a robust solution to this, though.
[1] https://code.visualstudio.com/docs/remote/containers
Has something like this ever been possible with Common Lisp and say Emacs?
Proc macros can run arbitrary code, so this POC is not that interesting - apart from raising awareness for the problem.
This can be done even easier without users having to use a macro: with `build.rs` build scripts, which are run by default. So all you'd need is to compromise some popular dependency with a custom build.rs
Many other languages have the same (or at least similar) problem (Makefiles, npm hooks, ...)
There is an interesting proposal and prototype for compiling proc macros to WASM so they can be run in a sandbox: https://github.com/dtolnay/watt
But in the end it doesn't make that much difference: nothing prevents a random library from just reading your secrets and calling curl to send it to a server at runtime.
Build time execution is definitely an additional attack vector.
But if you use a third party dependency, you have to trust it or review all it's code for every version. There is no way around this, and it's true for any language.
This can be done even easier without users having to use a macro: with `build.rs` build scripts, which are run by default. So all you'd need is to compromise some popular dependency with a custom build.rs
Many other languages have the same (or at least similar) problem (Makefiles, npm hooks, ...)
There is an interesting proposal and prototype for compiling proc macros to WASM so they can be run in a sandbox: https://github.com/dtolnay/watt
But in the end it doesn't make that much difference: nothing prevents a random library from just reading your secrets and calling curl to send it to a server at runtime.
Build time execution is definitely an additional attack vector.
But if you use a third party dependency, you have to trust it or review all it's code for every version. There is no way around this, and it's true for any language.
I would just like to tack on that malicious code is against the crates.io terms of service, and something like exfiltrating secrets in a build script is something that very clearly qualifies as malicious. If you ever encounter this in the wild, please make sure you report it to the crates.io team, so it can be removed.
I think it would be better to report here, https://rustsec.org/, and folks running cargo audit would be aware of the issue even if they’ve already downloaded the dependency.
Also, it’s not a new problem; a Makefile or configure script can run arbitrary code as well.
Yes this has always made me wonder about the pushback you see with the recent move towards curl | sh installers. In the past you'd download a random tarball and then run ./configure which could do anything.
There are three main problems with curl | sh: the file one the web server could be replaced without modifying the source in version control (and unlike a git checkout, the hash of the file is not verified), you can’t read the code before it runs, and curl could fail to download the whole file.
Of course, I bet a lot of people don’t bother to read any of the source code of a program that they’ve downloaded anyway.
Of course, I bet a lot of people don’t bother to read any of the source code of a program that they’ve downloaded anyway.
Downloading a tarball and running ./configure from it (pretty dang common) also does not have the changes checked into version control, nor the hash verified.
Same is true of `npm install`, deb/rpm/etc packages, etc: you don't have proof what was distributed to you matches up with what was in VCS.
You can read the code before it runs and solve the "curl could fail" theoretical arguments by just.. removing `| sh` and examining + running yourself.
Same is true of `npm install`, deb/rpm/etc packages, etc: you don't have proof what was distributed to you matches up with what was in VCS.
You can read the code before it runs and solve the "curl could fail" theoretical arguments by just.. removing `| sh` and examining + running yourself.
I agree; distribution via git is better in many ways than distribution via tarball. I believe that npm and similar package managers mostly pull code from git repositories. Of course, even then you might want to double check that the package name hasn’t been hijacked or sold off.
Of course you can break the curl|sh into separate steps and check that the script isn’t malicious before you run it, but the fact that you have to do that makes it a bad idea to distribute software this way. If you were told to download an installation script, inspect it, and only then to run it then there would be less of a problem. curl|sh is yet another sign that we so often prefer convenience over reliability and safety.
Of course you can break the curl|sh into separate steps and check that the script isn’t malicious before you run it, but the fact that you have to do that makes it a bad idea to distribute software this way. If you were told to download an installation script, inspect it, and only then to run it then there would be less of a problem. curl|sh is yet another sign that we so often prefer convenience over reliability and safety.
Debian are working on reproduceable builds for apt. Not there yet but going in the right direction.
https://wiki.debian.org/ReproducibleBuilds
https://wiki.debian.org/ReproducibleBuilds
You can break the pipe and curl the file first, read it, and then run it. But I doubt that anyone ever reads through the thousands of lines of m4 that come with a typical program that uses autoconf either.
If the project uses Autoconf/Automake, then you can just read the .in files instead. If they include anything unexpected, then it will be pretty obvious (since anything unexpected will be a lot more complicated–looking than anything normal). But if they do include a bunch of custom m4 files, then you’re going to be spending more time on it than you would want.
You can solve the third problem by declaring a shell function `install` that is run at the send of the script. The first problem is a problem but, as far as I know, most language package managers don’t verify provenance anyway: yarn install foo can perform arbitrary side-effects either directly or through its transitive dependencies.
Citation needed. Show me a Makefile + IDE combination that executed code by simply opening a file. I think you’re missing the language server part of this.
IntelliJ with "Build in background" enabled?
https://www.jetbrains.com/help/idea/executing-build-file-in-...
https://www.jetbrains.com/help/idea/executing-build-file-in-...
bash autocompletion will run arbitrary code from a Makefile. I wouldn't be surprised if many editors do too.
I think programmers' editors in general have treated "automatically run arbitrary code supplied by files you're editing" seriously as a security vulnerability since sometime around 2000.
(For example, Emacs realised that 'local eval' wasn't a good thing to have enabled globally in Emacs 19, in 1994, and spent the next decade or more closing many other loopholes involving local variables specified directly in files.)
If modern editors and IDEs are no longer thinking that way, I think that's a mistake.
(For example, Emacs realised that 'local eval' wasn't a good thing to have enabled globally in Emacs 19, in 1994, and spent the next decade or more closing many other loopholes involving local variables specified directly in files.)
If modern editors and IDEs are no longer thinking that way, I think that's a mistake.
Does opening an android project in android studio implicitly run any of the code in the project? I'm guessing it does, because the ide seems to be very busy all the time, even when idle
build.gradle can contain arbitrary Groovy code with full system access, and needs to be executed to figure out the project structure.
Apart from running make which of course runs the makefile, under what scenario does viewing a makefile run it?
Makefiles can actually be quite dynamic, so a program merely trying to figure out the list of target has to execute code. For example put this in a Makefile and do `make <TAB>`, the file will be created (no need to press enter):
VALUE := $(shell touch /tmp/something)FWIW, while both bash and fish completion execute "make -n" for tab completion that isn't the case for zsh. zsh uses an internal parser for makefiles¹, and as such won't execute the shell function or recipes that use the + prefix.
¹ https://github.com/zsh-users/zsh/blob/master/Completion/Unix...
¹ https://github.com/zsh-users/zsh/blob/master/Completion/Unix...
You're not just viewing it. You're opening it in an IDE which compiles it behind the scenes for you.
Many IDEs also do this for other languages (e.g. by running make), and the same problem applies.
Many IDEs also do this for other languages (e.g. by running make), and the same problem applies.
> But in the end it doesn't make that much difference: nothing prevents a random library from just reading your secrets and calling curl to send it to a server at runtime.
The difference here is that it happens when you open the project in the editor. If I'm suspicious of some code my first reaction would be to open it my editor and inspect it.
The ESLint extension always asks whether you trust the `eslint` executable before it's enabled. It's still quite easy to click "allow" without thinking about it, but at least you'll have a choice to not execute potentially random code.
The difference here is that it happens when you open the project in the editor. If I'm suspicious of some code my first reaction would be to open it my editor and inspect it.
The ESLint extension always asks whether you trust the `eslint` executable before it's enabled. It's still quite easy to click "allow" without thinking about it, but at least you'll have a choice to not execute potentially random code.
That's a really recent feature and I'm sure Rust-analyzer will support it soon.
I suspect the same problem exists in many other languages. How can you open a CMake project without executing it?
I suspect the same problem exists in many other languages. How can you open a CMake project without executing it?
In a text editor....
> Many other languages have the same problem (Makefiles, npm hooks, ...)
This simply isn’t true. All of these require an action by a user to execute the command (e.g npm install, make build). What the author is claiming is that a typical rust LSP setup will execute the arbitrary macro code simply by viewing the file in certain IDEs.
Feel free to show me an example of this in makefiles or npm and I’m happy to retract.
This simply isn’t true. All of these require an action by a user to execute the command (e.g npm install, make build). What the author is claiming is that a typical rust LSP setup will execute the arbitrary macro code simply by viewing the file in certain IDEs.
Feel free to show me an example of this in makefiles or npm and I’m happy to retract.
Visual Studio has the same problem with C#. By default VS will load all defined analyzers into Roslyn which can execute arbitrary code.
VS warns you with a confirmation dialog that shouldn't just be ignored because "I just want to look and not compile". So, don't open any random .csproj or .sln and assume you are safe.
VS warns you with a confirmation dialog that shouldn't just be ignored because "I just want to look and not compile". So, don't open any random .csproj or .sln and assume you are safe.
There aren't a bunch of languages with proc-macros and IDEs. That'd be where you'll see a major intersection. (Maybe C++ has this problem with some ides?)
Languages with similar risks are ones where a Repl is is the key form of development. In those scenarios you are also one bad dependency from stolen info.
Languages with similar risks are ones where a Repl is is the key form of development. In those scenarios you are also one bad dependency from stolen info.
That's not really relevant though. Anything that runs code on my computer without my awareness of it should be considered a security bug.
Alas, the nature of computation makes this only ever a matter of squinting hard enough at the problem.
Just as it turns out that matter and energy are almost the same thing seen from a different point of view, it's the same with code and data. Running code and processing data are no different to a computer.
You think a picture of a dog and a Windows program are plainly different kinds of things, the computer does not agree.
Something like Wuffs † aims to at least control the blast radius. If (in some alternate or far future world) you were only ever looking at pictures of a dog via Wuffs, you could at least feel confident that doing so did not have some entirely unforeseen consequences, like exfiltrating your SSH private keys. Today you certainly can't be sure of that, none of the tools you use have such a cautious approach.
† https://github.com/google/wuffs
Just as it turns out that matter and energy are almost the same thing seen from a different point of view, it's the same with code and data. Running code and processing data are no different to a computer.
You think a picture of a dog and a Windows program are plainly different kinds of things, the computer does not agree.
Something like Wuffs † aims to at least control the blast radius. If (in some alternate or far future world) you were only ever looking at pictures of a dog via Wuffs, you could at least feel confident that doing so did not have some entirely unforeseen consequences, like exfiltrating your SSH private keys. Today you certainly can't be sure of that, none of the tools you use have such a cautious approach.
† https://github.com/google/wuffs
You can also just put arbitrary code in build.rs, it will be run by cargo check, rust-analyzer, etc. Though I admit macro expansion hacks are more fun and easier to hide.
For what it's worth, any VSCode extension that integrates with language tooling could be used to implement this.
This is an inherent problem of languages where execution is needed to understand its semantics. Most interpreted languages have this issue, and Rust has this issue due to proc-macros being Rust code that needs to be compiled and executed to process other Rust code.
The more general problem with trusting software supply chains is well described in Ken Thompson's Turing award lecture on "Trusting Trust"[1]
[1]: http://users.ece.cmu.edu/~ganger/712.fall02/papers/p761-thom...
[1]: http://users.ece.cmu.edu/~ganger/712.fall02/papers/p761-thom...
This doesn't affect me because I https://www.grepular.com/Sandbox_Rust_Development_with_Rust_...
Blocks of text again... TL;DR?
as linked in the readme:
https://www.youtube.com/watch?v=RRLw3OBJ0fM
https://www.youtube.com/watch?v=RRLw3OBJ0fM
There's a TL;DR at the top
This is why VSCode is adding Workspace Trust to prevent extensions from running untrusted code by merely opening a directory.
https://github.com/microsoft/vscode/issues/120251#issuecomme...
Are there any working groups or teams in the rust foundation[0] looking into stuff like this? I know every package manager has these issues but there's no technical reason preventing us from building sandboxes (i.e. WASM, deno, ...) for this and making it a first class citizen of cargo/rustup/etc.
Just installing a relatively popular crate (say Hyper) makes you realize that all of your secret could have been stolen by any of the myriad of dependencies.
[0] https://www.rust-lang.org/governance
Just installing a relatively popular crate (say Hyper) makes you realize that all of your secret could have been stolen by any of the myriad of dependencies.
[0] https://www.rust-lang.org/governance
Well, the topic has come up on the internals forum from time to time, e.g.:
https://internals.rust-lang.org/t/pre-rfc-procmacros-impleme...
I don’t think there’s an active working group though.
https://internals.rust-lang.org/t/pre-rfc-procmacros-impleme...
I don’t think there’s an active working group though.
Is there a reason that access to the filesystem isn't sandboxed aggressively by the compiler? Even having build macros that can access arbitrary parts of the filesystem (vs a dedicated scratch directory) seems like a bad idea. Is there any legitimate use-case here?
rust-embed is one: https://docs.rs/rust-embed/5.9.0/rust_embed/trait.RustEmbed....
This macro lets you embed an entire folder of assets in your binary at compile time, to simplify distribution.
Taking the concept further, I could also imagine build macros that compile Typescript or SASS files at build time, or generate data structures from a Protocol Buffers definition file, or in general operations that ingest non-Rust source code and use tools outside the repository.
This macro lets you embed an entire folder of assets in your binary at compile time, to simplify distribution.
Taking the concept further, I could also imagine build macros that compile Typescript or SASS files at build time, or generate data structures from a Protocol Buffers definition file, or in general operations that ingest non-Rust source code and use tools outside the repository.
Sure. I would expect such a tool to be happy enough with sandboxing to the folder containing the source of the project and the build folder, no?
Meh. You could do this in C also. Nothing new here.
The PoC doesn't even open a file, it just opens the directory. It's a pretty big difference, when you execute a build script you _expect_ to run code, when you open a directory in your editor you don't expect any side effect _at all_.
My guess is that since the proc_macros returns a TokenStream, rust-analyzer have no way to know what it provides except running it.
I'm not sure there's a solution for this that doesn't cripple macros in Rust, apart from being able to configure rust-analyzer to ignore the macros, which clearly limit its usefulness.