Sounds to me like Cursor internally has a private NPM registry with those packages. Because of how NPM works, it's quite easy to trick it to fetch the packages from the public registry instead, which could be used by an attacker [0].
Assumably, this Snyk employee either found or suspected that some part of Cursor's build is misconfigured as above, and uploaded those packages as a POC. (Given the package description "for Cursor", I'd think they were hired for this purpose.)
If that's the case, then there's not much to see here. The security researcher couldn't have used a private NPM registry to perform the POC if the point is to demonstrate a misconfiguration which skips the private registry.
.
[0] In particular, many proxies will choose the public over the private registry if the latest package version is higher: https://snyk.io/blog/detect-prevent-dependency-confusion-att...
we did not hire snyk, but we reached out to them after seeing this and they apologized. we did not get any confirmation of what exactly they were trying to do here (but i think your explanation that someone there suspected a dependency confusion vulnerability is plausible. though it's pretty irresponsible imo to do that on public npm and actually sending up the env variables)
Wouldn't it be more like "pretty illegal"? They could have simply used body: JSON.stringify("worked"), i.e. not sent target machines’ actual environment variables, including keys.
They could have demonstrated the POC without sending data about the installing host, including all your environment variables, upstream. That seems like crossing the line
Allowing someone full access to the contents of your environment (i.e. output of env command) is a big deal to most, I suspect.
What's the justification for taking all of the environment variables? This post tries to paper over that particular problem. If your goal was to see if you could attack the dependency chain the first steps of user+hostname would have been sufficient to prove your case.
Taking the environment variables is about taking the secrets, and kind of moves this from PoC to opposition supply chain attack. Not to mention it's not only Cursor devs that would be affected by this, it could have (if your plan worked) attacked anyone using the extensions.
It's also a tough buy given the note about the Snyk cofounder moving to compete directly with Cursor (courtesy @tankster): https://techcrunch.com/2024/11/14/tessl-raises-125m-at-at-50...
Assuming truly innocent motivations, you guys still need to give your heads a shake and rethink your approaches here.
Who did the GDPR review before extracting env vars from systems that were not under your control? How did actively extracting potentially private data from the environment not get flagged as Unauthorized Access?
If this "experiment" (which happened to be against a competitor, mind) was reviewed and approved internally, that is a great demonstration of Snyk's approach to (ir)responsible data collection and disclosure.
I hope there is no foul play.
IDEs, plugins, development utilities, language libraries, OS packages, etc. So much code that I take on blind faith.
Several years ago I worked somewhere that prohibited web browsers and development tools on laptops. If you needed to use a browser, you’d have to use one over Citrix. If you needed to code, you’d use a VDI or run the tools in a VM.
At the time I thought their approach was clinically insane, but I’m slowly starting to appreciate it.
Unlike with Terraform/OpenBao, I know of no community effort effort to keep the open-source version of this project alive. The latest open source version is still available on the Ubuntu repo, but who knows who long it will work until somefor of bit rot occurs.
Except the vm is some old windows version without any tools on it. no shell access.
can't actually do anything useful on there at all.
VDI systems could work if implemented properly. but that's the last thing a security team actually wants to do.
A newer project fires up VMs from a Python script that calls an adapter for EC2 (with the boto library) when run on AWs and for VirtualBox (by calling VBoxManage) when running locally. That allows us to simulate EC2 locally: it's a project that has to deal with many long jobs so we start VMs for them and terminate the VMs when the jobs are done. That also runs better on our mix of development systems. WSL2 helped to ease the pains of developing on Windows. We call the native Windows VirtualBox, not the one we could have installed inside WSL2, but we keep most of the code that runs on Linux.
None of my jobs since have had any semblance of a better way to set up a local dev environment easily.
It was just way easier to encapsulate services or other things in a quickly reproducible state.
I digress..
Let’s be clear, it’s still clinically insane, even if marginally rationalized.
A hybrid(?) alternative is enroot, which is pretty neat IMO, it converts a docker container into a squashfs file that can be mounted rw or used in an ephemeral way. https://github.com/NVIDIA/enroot
nvidia gates it's virtualized GPU offerings behind their enterprise cards, so we're left with ineffective command translation.
IMO: I can tolerate just about every other type of VM overhead, but choppy/unresponsive GUIs have a surprisingly bad ergonomic effect (and somehow leak into the performance of everything else).
If we could get that fixed, at least amongst Linux-on-Linux virtualization, I think virtualizing everything would be a much more tenable option.
If you can't do that because it's for company/corporate purposes then I can sympathise with not wanting to pay Nvidia's prices.
Specifically I'm transitioning my Go and Zig development environments from an old mac to an M1 with Asahi Linux and getting a bit lost even finding replacements for Truecrypt and Little Snitch. Do these VM tools support encrypted VM's with firewall rules? I saw Vagrant mentioned here and that sounds like it might cover the network isolation, but what else would you suggest?
LXD, unlike Docker, doesn't play fast-and-loose with security. It runs rootless by default, and I don't allow non-root access to the LXD socket on host. Each container is a full userspace, so it's much more convenient to configure and use than Dockerfiles.
SSH from a container to a remote works transparently because I forward my SSH Agent. This is secure because my agent confirms each signing request with a GUI popup (on host).
These days I work in JavaScript and rarely have issues with project environments interfering with each other. I've gotten lazy and don't use VMs anymore.
In theory docker type setups could work but they just seem so much effort to learn and setup.
The VM might protect me, but it will not protect the users of the software I am producing. How can I ship a product to the customer and expect them to safely use it without protection when I myself only touch it when in a hazmat suit?
No, that is not the environment I want.
My current solution is to be super picky with my dependencies. More specifically I hold the opinion that we should neither trust projects nor companies but only people. This is not easy to do, but I do not see a better alternative as for now.
i've got a simple per-project dot file that describes the file system binds. while i'm working on a project, new terminals i open are automatically isolated to that project based on that dot file. it has very low (cognitive) overhead and integrates pretty much seamlessly. i suspect many developers have similar scripts. i looked for projects that did this some time ago, but couldn't find it. either because it's too simple to make a project about, or because i don't know how others would describe it. if anyone has pointers...
i don't limit network access (though i did experiment with logging all traffic, and automatically setting up a mitm proxy for all traffic; it wasn't convenient enough to use as regular user). there is still a whole kernel attack surface of course. though i'm mostly concerned about files being read/destroyed.
If you ran a key logger on my machine you would never get into any major site with mfa. You couldn't watch me log on to the azure console with passkey and do much with it. But if you scrape a saved key with publish abilities bad things happen.
I did not start doing this because of security, but just to get something more or less self managed without any possibility to break different projects. I am tired of my team spending too much time on extensions, versions, packages, ...
Docker compose files have saved our team many hours, even if it's extremely wasteful to have multiple vscode instances running alongside each other
For work use I use a work machine and if it gets compromised it's not really my own problem.
Is that really a good mindset for a organization?
It doesn't fully sort out the trust issue though, even if everything is sandboxed in some fashion.
Also, many people here work on multiple projects for different customers. Having a security breach for one affecting the other is not something you'd be happy with.
> But in general, it’s a good idea not to install NPM packages blindly. If you know what to look for, there are definite signals that these packages are dodgy. All of these packages have just two files: package.json and index.js (or main.js). This is one of several flags that you can use to determine if a package is legit or not.
This works -- maybe OK for top-level packages. But for transitive dependencies it's nearly impossible to vet every transitive dependency.
If you're pulling in a package that has 400 dependencies, how the heck would you even competently check 10% of that surface area? https://gist.github.com/anvaka/8e8fa57c7ee1350e3491#top-1000...
This would be where different security advice would apply: don't pull in a package that has 400 dependencies.
im not sure what has to change in UX to make these approaches more palatable, but if you have to frequently allow 'good' behaviors, my experience is it never takes off.
It's almost trite at this point to comment on the obsession that Node has created with developers to reduce functionality to the smallest possible reusable parts, even trivial things, and publish them as packages, then to import and use those dependencies. The idea, in and of itself, is not terrible, but it's been taken to a logical extreme which is very much the definition of terrible.
Going through that list... they all look like personal projects, with no dependents, and a single release by a single person.
Then because there is no standard library, many reinventions of similar but incompatible utilities. etc.
It's very much based on reality. The npm ecosystem is just absolutely fucked.
At my place of work we use this great security too called Snyk. Definitely check it out
/s
They also mark projects as "abandoned" if they move to any other forge that isn't github. And they stay abandoned even if new releases appear on npm/pypi :D
Their competence isn't as big as their fame, in my opinion.
Also one of their sales people insulted me over email, because apparently not being interested in buying their product means you're an incompetent developer who can only write software filled with vulnerabilities.
Completely backwards software that corpos only seem to buy because their insurers force them to check off some security list box.
edit: or microsoft outlook sucks… I tried to sort in reverse my inbox to see what's the oldest email there and "the request cannot be satisfied"
Codeberg looks interesting, and there are self-hosted ones like Forejo that also look great if you're okay with the maintenance.
It has CI, pull requests, issues and whatnot. It also doesn't force you to use 2fa if you don't want :D
If you do corporate open source though, you're stuck on github because snyk, openssf, pypi and whatnot only interface with github.
For actual libre software codeberg is very good.
Keep in mind that debian salsa is open to everyone as well. The only annoyance is that non debian developers have a "-guest" suffix. But it's ok to use for projects that aren't debian specific.
Well theres a sign of a good team.. /s
That's actually an interesting take, I haven't heard too much about them except that they do have an ego.
You're not supposed to leave public artifacts or test on public services during an assessment.
It's possible Cursor asked them to do so, but there's no public indication of this either. That's why I qualified my original comment. However, even if they did ask them to, it's typically not appropriate to use a separate unrelated public service (NPM) to perform the demo.
Source: I've done a handful of security assessments of public packaging indices.
They should be running a private npm repo for tests (not difficult to override locally) and also their own collaborator server.
I think what we're more likely to see is that leglislation like DORA and NSIS increasinly require that you audit third party packages. This enforcing a different way of doing development in critical industries. I also think you're going to see a lot less usage of external packages in the age of LLM's. Because why would you pull an external package to generate something like your OpenAPI specification when any LLM can write a cli script that does it for you in an hour or two of configuring it to your needs? Similarily, you don't need to use LLM's directly to auto-generate "boring" parts of your code, you can have them build cli tools which does it. That way you're not relying on outside factors, and while I can almost guarantee that these cli tools will be horrible cowboy code, their output will be what you refine the tools to make.
With languages like Go pushing everything you need in their standard packages, you're looking at a world where you can do a lot of things with nothing but the standard library very easily.
I wouldn't install it if you paid me to, because it feels a lot like Unit 8200 pumps out entrepreneurs and funds them so that (like the NSA) they have their foot already in the door.
It's the equivalent of observing that SFBA startups tend to have a lot of Stanford grads at the helm.
(I don't have any particular love for Snyk as a product suite. I think most supply chain security products are severely over-hyped.)
For example, if you're using a multi-protocol networking library, and it says that the version you have installed is has a vulnerability in its SMTP handling, but you don't use the SMTP functionality, is that a FP?
I'd argue that it's irrelevant, but not a false positive.
I never had it get the version of a library wrong.
Snyk does follow a responsible disclosure policy and while no one picked this package up, had anyone done so, we would have immediately followed up with them.
In response, you suggest that you'll send a letter of apology to the funeral home of anyone that got hit. Compromising their credentials, even if you have "good intentions", still puts them into a compromised position and they have to react the same as they would for any other malevolent attacker.
This is so close to "malicious" that it's hard to perceive a difference.
edit: Let's also remind everyone that a Snyk stakeholder is currently attempting to launch a Cursor competitor, so assuming good intentions is even MORE of a stretch.
https://snyk.io/blog/snyk-security-labs-testing-update-curso...
But all of that is why they get popular. Lazy, crappy, easy things are more popular than intentional, complex, harder things. Shitty popular tech wins.
Npm doesn't really do namespaces. There's just no ownership to prove as most packages are published like "call-home" with no namespace required. This gives exciting opportunities for you to register cal-home to trap users who miss type, or caII-home to innocuously add to your own or open source projects or whatever. Fun isn't it?
In this case the call home package is namespaced, but the real attack is the packages like "cursor-always-local" which has no namespace. Which can sometimes (?) take precedence over a private package with the same name.
It's not a pretty picture, you were better off missing it really.
Yes it really does. npm has namespaces (called scoped packages) and even explicitly encourages their use for private packages to avoid this sort of attack. From the npm docs: "A variant of this attack is when a public package is registered with the same name of a private package that an organization is using. We strongly encourage using scoped packages to ensure that a private package isn’t being substituted with one from the public registry." [1]
> This gives exciting opportunities for you to register cal-home to trap users who miss type, or caII-home to innocuously add to your own or open source projects or whatever. Fun isn't it?
npm actively blocks typo-squatting attacks during the publishing process: "Attackers may attempt to trick others into installing a malicious package by registering a package with a similar name to a popular package, in hopes that people will mistype or otherwise confuse the two. npm is able to detect typosquat attacks and block the publishing of these packages." [1]
This thread is full of people demonstrating the concept of confirmation bias.
Stopped using it for serious stuff after I noticed their LLMs grabs your whole .env files and sends them to their server... even after you add them to their .cursorignore file. Bizarre stuff.
Now imagine a bad actor exploiting this... recipe for disaster.
If you want secure LLM you can use Mistral, which comes with all the EU limitations, good and bad.
Wouldn't a lot of small packages consist of just these two files, meaning seeing just these two files in a package may raise an eyebrow but hardly be a smoking gun?
You have any idea how easy it is to fake it?
Also, snyk doesn't scan code that isn't on github, because they are under the impression that all the code in the world is on github, so things like gnome.org, debian salsa or codeberg are completely ignored.
So you won't get reliable data from snyk.
edit: snyk doesn't scan code at all, they rely on unrelated "metrics" to give a rating that is not very useful.
Helps in cases like this.
So any third party code changes end up in git commits and are easily visible and reviweable.
So running npm update/upgrade includes the code that changed in the dependencies in the commit.
seems to be either a tool that isnt out yet or perhaps not available for free or the public.
Yes, but also impractical.
The TL;DR is that our security research team routinely hunts for various vulnerabilities in tools developers use. In this particular case, we looked at a potential dependency confusion attack in Cursor, but found no vulnerabilities.
There's no malicious intent or action here, but I can certainly understand how it appears when there's not a ton of information and things like this occur! As a sidenote, I use Cursor all the time and love it <3
And exfiltration was needed to confirm a vulnerability why exactly?
I love how completely unaware you guys are.
Ethically, your work was even lower than that of those who test their AI tools on FOSS code, send in bogus reports and thus waste maintainer's time. Experimenting on unwitting humans and ecosystems is not okay.
I'm fairly concerned about the state of Python packages. It's not every week, but I frequently stumble upon packages that are not what they appear to be. Sometimes not maliciously, sometimes the author just got overly ambitious and failed to deliver, other times, someone is clearly typo-squatting or attempting to get you to install the wrong thing.
Think xz-utils but even much less sophisticated exploits.
I don't see any systematic protection against this?
Your criticism sounds to me like "just a reminder that this armed bodyguard service comprises Navy SEALs and Army Rangers". Uh, great!
Most companies where I've worked as a security researcher, you get some time as part of your job to hack on random stuff to be able to generate interesting talks / research. This feels like that.
This isn't a special cyber spooky practice, most pentesting companies do this to generate IP (rarely, lol), buzz (reasonably often) and keep the staff happy (this is really the main thing).
It's rare for management to be fully across the scope of this.
It has less to do with whether it malicious intent from the start of building an organization for explicit intent of capturing core infra. It has more to do with how the Government of Israel operates and the legal requests they can make of their citizens and/or veterans.
Perhaps concern over Israeli products should be probably higher than for China as Israel more well known incidents of exploits as a State actor like with Stuxnet, Pegasus or more recently with pagers etc.
China no doubt has their own share of operations but they either have not used them as publicly in a large scale overt operation or been more discreet about it.
Point is the concern is valid just as it would be valid for China.
In other words, it's staffed by Israeli kids who made the choice most of us would have made under the circumstances. It seems a bit unfair to hold that against them more than 10 years later, no?
You could say the same about the guy in a call center in India trying to pull a tech support scam on you over the phone. Yes, he's probably making the best choice he can for his own livelihood, probably the same thing you would do in his position. No, that doesn't mean you should trust him.
Just as you have to treat all Chinese companies as under control of the PRC government and all Australian companies as compromised by their security services, you have to treat all Israeli citizens as under the control of the Israeli military. Any adult can be conscripted and they have a history of disguising military operations as civilian ones. Someone might tell you they left the Israeli military 10 years ago and they're probably telling the truth, but if you make a habit of believing that statement you're going to get burned.
Did Cursor made claims to this effect and invited public to hack them?
Or are you equating someone saying they "take security seriously" to "it's an open season, please attack our systems."?
I was dismayed to learn about their choice of brand, and think it might cause confusion. :(
