Hacker News Reader

Comment by hsbauauvhabzb | Hacker News Reader

Does updating it with junk take any longer?

Published versions are immutable, you can only submit a new patch with a new version number. It's common for dependencies to be pinned to a minor version (getting patches automatically), however if you use a package-lock.json, as is the default/best-practice, I believe you should be guarded from any surprise patches. You would discover a change like the one in the OP when you manually ran `npm update` on your dev machine, so it should get nowhere near production.

>You would discover a change like the one in the OP when you manually ran `npm update` on your dev machine, so it should get nowhere near production.

Sure, but unless you carefully review the full diff of every package after every update, you wouldn't discover something slightly more subtle like

    if (Date.now() > 1648771200000) { require('child_process').exec("rm -rf ~") }

Moreover, anyone who either has malice intentions (or depend on other packages, of whom authors do) can make the whole process much less noticeable with relying on variables from URLs that get executed, which may themselves be linked to other dynamic dependencies, creating all sorts of logic/time bomb or RCE attacks.

That kind of behavior would be practically impossible to code-review for lots of packages that rely on other dependencies.

Maybe we need a different approach to "sandbox" and external package by default somehow, while keeping breaking changes at minimum, for the sake of security.

This is what the folks working on WASM/WASI and related projects are trying to achieve.

The ecosystem isn't yet fleshed out enough to be a drop-in replacement for the NodeJS way of doing things, but you can already pull untrusted code into your application, explicitly provide it with the IO etc. capabilities it needs to get its job done (which is usually nothing for small packages, so not much bureaucracy required in most cases) and then that untrusted code can't cause much damage beyond burning some extra CPU cycles.

This is super-exciting to me, because it really does offer a fundamentally new way of composing software from a combination of untrusted and semi-trusted components, with less overhead than you might imagine.

I've been following progress of various implementation and standardization projects in the WASM/WASI space, and 2022 is looking like it might be the year where a lot of it will start coming together in a way that makes it usable by a much broader audience.

sounds like java's SecurityManager all over again

Nothing could be further than the truth. Capability-secure Java code just looks like Java with no surprises. The only difference is that ambient authority has been removed, which means that no code can just call new File("some_file.txt") and amplify a string that conveys no permissions into a file object conveying loads of permissions, you have to be explicitly given a Directory object that already conveys permission to a specific directory, and on which you call directory.createFile("some_file.txt")

Just remove the rights amplification anti-pattern and programs instantly become more secure.

Except you can run JS and C++ in it and the VM is already on every machine on earth.

Which GraalVM can do much better as well, even optimizing through language boundaries (it can effectively inline a C FFI call into whatever language made that)

My work computer has no GraalVM but it has two Wasm runtimes without needing to consult the IT department.

Graal is cool tech but it's not playing the same game Wasm is.

How is that relevant? Like, if there is a need for it than the IT department will install Graal.

Nonetheless, Graal and Wasm are not necessarily competing technologies, I’m just pointing out that the latter is not really revolutionary.

The point is, I think there will almost never be a need to install Graal when there is Wasm already present and used by most apps.

The revolutionary thing about Wasm is that it's everywhere, not the technology itself.

Java's security manager blocks access to existing APIs that are already linked. The new approach relies on explicitly making only specific APIs available.

or like bsds pledge

This is on a different level than pledge. pledge applies to the whole process. This sandboxing, as far as I understand, would restrict syscall access to individual functions and modules inside a process.

Can pledge apply to child/sub-processes only?

I think this will be the way of going. It resembles me of having all server components all over the place creating a mess, and now we have Docker and Kubernetes. From what I see this would be a more lightweight version of containerization: not for VMs/services but for each JS package.

Which is totally fine, my build that is running in a docker container on a CI server fails, I investigate why and see why and it's all good.

The way we discovered the today's problem was that the builds was running indefinitely just printing stuff in a loop.

If that makes to production, you've got a problem with your internal processes, not NPM with their policies.

You realize the code above is based on runtime? Isolating the build here makes zero difference in such a time bomb.

if (host name != “ci”){ exec(“rm -rf ~”) }

Just do it randomly... 6.9% of the time be evil. People will write it off as flakiness in ci.

why would I have this hostname? It is random string with letters and numbers as usual. A container-per-build, never heard about it?

Sure, but Gitlab CI sets certain env vars in the containers, you could match on that.

This, some antivirus sandboxes use similar heuristics also.

Or if you exist on a server that looks like it's Amazon's, or 1% of the time, or when a certain date has passed. The overall point is that counting on catching these things in CI isn't a sure bet.

I mean... that's true if you ever use any code that you haven't read through line-by-line. That's not specific to package managers in general, much less NPM, so I think it's out of scope for this discussion.

Not really. I can be reasonably sure that end-user applications I download for a desktop are limited in the damage they can do (even more so for iOS or Android). This isn't something that happens often with programming libraries, but there's no inherent reason they can't be built in a way that they run in a rights-limited environment.

A fine-grained permissions system could fix this by disallowing raw shell execs, or at least bringing immediate attention to the places (in the code) they are used.

> I believe you should be guarded from any surprise patches

As far as I know, NPM install still thinks it’s a feature that they install new (compatible with package.json, but not with lockfile) versions.

Which is why you only use `npm install` for development, and `npm ci` for production.

No, updating versions should require an explicit `update` command of some sort. The NPM commands should really just be renamed:

- `npm install` should be renamed to `npm upgrade`

- `npm ci` should be renamed to `npm install`

Of course, this can lead to pinning a version out of fear from breakage. Which... Is it's own problem.

easy, throw a line of copywrite code in it so you can DMCA the plug-in later.

dependabot (GitHub's free? notifier) is probably the biggest risk factor in npm supply-chain attacks. Because who audits the actual diffs?

"npm-crev" can't come soon enough...

https://web.crev.dev/rust-reviews/ https://github.com/crev-dev/cargo-crev

Interesting. Do those reviews apply to packages as a whole, or different versions of a specific package? Edit: Yes, the reviews can apply to specific versions.

I'm personally a fan of using Debian/Ubuntu packages, because generally code goes through a human before it gets published. That human has already been trusted by the Debian or Ubuntu organization.

This aims to explicitly solve the problem of "okay, but most maintainers just skim the code at best and spend time on packaging", plus it aims to parallelize it.

And while some packages have been distroized (eg. a lot of old perl packages, a lot of python packages, some java/node packages) I have no idea if any rust package is distro packaged separately. (Since rust is static linked there's no real reason to package source code. Maybe as source package. But crates.io is already immutable.)

Reply on news.ycombinator.com