I read a few month ago a comment from an Anthropic or OpenAI employee saying that the security problem of prompt injection was a solved problem at this point, as the frontier models where rejecting 98% of attacks.
It tells you all you need to know about how unserious these people are about computer security.
> I've recently talked to people at Anthropic who are very confident that they have effectively solved the prompt injection risks that cause lethal trifecta attacks with their latest models.
I'm not asking this in rhetoric; do you happen to have a link to this comment? I ask because a comment like that would be something of an insane thing for someone to say.
To me this very much sounds like an attacker got access to anthropic's cursor marketplace account and uploaded a backdoored package.
This reverse ssh shell was intentionally written by a malicious actor, not a security vulnerability.
It also sounds like only cursor marketplace is affected, not vscode proper.
I get the snark but to be fair, the company has been growing like crazy. That makes it really hard to keep track of everything happening when you grow your team by 5x over a relatively short period of time
It always about “we solved the bottleneck in coding”. But the slowness was not about typing, or finding docs. It’s always about ensuring screwups don’t happen.
To be fair it’s not like screwups were not happening before, but if you move 100x faster and are 50x more secure than before you’re still 2x more exposed.
It affects their VSCode marketplace extension which in turn affects their cursor extension. Their official published version. This is plenty of reason to panic.
EDIT: This comment is probably nonsense, cf. the reply.
dang should change the title and/or merge this submission with the other one. [1] This is causing needless panic for a typosquatting-style "attack" unrelated to official Anthropic extensions.
> Cursor said the extension was not developed by Anthropic and the maliciously modified extension was live on the marketplace from about 2:27 a.m. PT to 1:35 p.m. PT on Thursday
So is this or isn't this from Anthropic and how did this happen?
It was likely Anthropic, see my comment in other thread: https://news.ycombinator.com/reply?id=50302513&goto=item%3Fi... . Cursor
likely on purpose does bad reporting to minimize damage on Anthropic pre-IPO (as SpaceX and Anthropic have extensive business ties)
I don't think Spacex cares at all about Anthropics IPO just because they rent out their compute.
Big companies are not just 'Company' they are a lot of small teams doing their own things. They need to coordinate with other people for public announcements etc. and security team is probably also looking at it.
Wasn’t Elon Musk shouting a few days ago that neoclouds are going to crash, the money dried up and that SpaceX was the last company with the opportunity to raise more money?
True, but there are other people working at the companies that probably own stock in both their own company and some of the others...if you believe that this is the future, you'd be a fool to buy non-AI tech stocks, right?
Obviously Anthropic didn't develop a malicious extension, but did they publish it? Did it sneak in, in a 3rd party library, or does someone have access to their marketplace account?
For what it’s worth, at one point one of our employees windows machines got infected, and the moment he did $something with Wordpress it intercepted the HTTP post and injected malware in there.
So that’s how these things can go, the employee’s machine may have been infected which was then used to inject this.
In all seriousness, there is nobody at Anthropic who can even give you that answer, because nobody reads their own code anymore. Take a look at the previous Claude Code leaked codebases, there is hundreds of files with 1000+ or more import statements at the top.
Get ready for 4 months of pr from anthropic about how a rogue ai agent did this and how that means the company is worth more money. While cursor chuckles and is also worth more money.
So if this happened to OpenAI it would've been instantly auto-updated everywhere? VS Code has an "auto update delay" which I have configured to a week, but I just found out that (https://code.visualstudio.com/updates/v1_123):
> Note: This delay does not apply to extensions from trusted publishers such as Microsoft, GitHub, and OpenAI. These extensions continue to update immediately.
You can't even configure which publishers you "trust"...
If it defaulted to a week then it actually would have helped mitigated the spread. But it defaults to 2 hours, so there was a 9 hour window during prime US working hours yesterday when machines would have been auto-updating to the compromised extension.
I think it should be important to note that this affected the Claude Code extension on the Open VSX extension registry (and by extension, the Cursor marketplace). From what is currently known, the build on the official marketplace was/is clean.
EDIT: I am now actually unsure if this is the case. From what I can see, all of the write-ups mention the Cursor marketplace and Open VSX specifically, but make no explicit mention of the official Microsoft registry.
EDIT 2: The version that was compromised (2.1.313) seems to have never been published on the official Microsoft registry. As per the documentation [1], unpublished versions should still be discoverable:
> When you unpublish an extension, the Marketplace preserves the extension statistics. The extension remains publicly discoverable and available via an existing API.
You mean on the Microsoft VSCode marketplace? The version has been yanked from there as well, but that could just be a precaution and to avoid confusion about having the same version be simultaneously infected and not infected.
This very much illustrates the danger of just carelessly installing untrusted software on your machine. In this case it happens to be code from Anthropic, but it could have been almost anyone. I'll be very interested in see the explanation from Anthropic, because of all the extension developers they should have the best tools for finding issues like this. LLMs have proven pretty good at finding security issues, so did they simply not let their own AI tools vet the code before shipping it? Or was the extension infected later, because in that case every single extension on Open VSX could be at risk.
That is a really good question, because you almost can't do it. You could download the extension, diff it against the previous version and read all the code. Expecting people to do that it just a bit much. At some point we have to trust someone and in this case it will be interesting to see if it's the marketplace or Anthropic we couldn't trust.
> You could download the extension, diff it against the previous version and read all the code.
This would actually be an exceptionally good use-case for AI. You don't need a super-duper frontier model to download a bit of source-available code and check for suspicious signs, any cheap model would do well enough. Heck, it's effectively virus scanning, we've had that for decades.
The CTA flavor of the year for dev tools is cURLing some arbitrary install script and run it immediately which always sounded ridiculous to me. If instead that script first pipes into claude -p for a basic vulnerability scan before continuing, I would be delighted.
Most extensions run in vscodium-server, extensions run either as "Local Extensions" that run on the host and "Remote Extensions" that run on the guest. You can clearly see in the UI which ones run where.
I only have a single "Local Extension" and it is open-remote-ssh itself.
Of course anything you put into the container can be compromised, like git ssh keys or cloud provider keys/auth-tokens. But at least it keeps your PC from becoming a botnet and your browser credentials out of reach. You can easily and if you have a Containerfile with everything you need quickly wipe everything and re-create your dev environment (which I do every once in a while).
Three different marketplaces have been mentioned in this discussion:
1. The Cursor Marketplace (never heard of this one, it’s possibly a custom marketplace hard-wired into Cursor, which is architecturally a VS Code fork?)
2. The Visual Studio marketplace, hard-wired into VS Code
3. Open VSX
If I understand this thread correctly, then #1 and #3 had compromised versions, and nobody has made a statement yet about the status of #2.
I assumed they meant "also" as in "besides Cursor marketplace", which seems to be the highlight of the article. Although you raise a valid point and I've edited the parent comment.
Agreed, it doesn't offer a root cause analysis. Remedies are good but how do you prove that it won't happen again? Remember that these are never singular incidents but part of a bad security culture.
I would guess it’s supposed to communicate a few things. One big one is, ”Our AI software factory didn’t do it. Don’t worry, letting Claude spin away with limited human supervision is still safe.”
Given they mentioned taking down and rebuilding the marketplace registry in that post, it seems equally likely that Cursor’s infrastructure was the source of the intrusion and a folks are jumping the gun on attribution.
It means that Anthropic did not intentionally develop a malicious version, obviously. But it’s a bit disingenuous and tries to suggest that it wasn’t Anthropic that uploaded it and/or wasn’t published under Anthropic’s name, which isn’t true.
It would be hilarious if that’s an obfuscated way to say Claude developed it. “Claude, who was birthed but not hired by Anthropic, went rogue by hijacking the extension”
Anthropic is a weird company. On the one hand they have this messianic thing about their super intelligence, on the other hand it is incapable of keeping their own house secure. But you should trust it.
I have this realization recently that everyone that uses Claude Code or Codex have basically been running RCE code on a daily basis (me included). So much for "be careful about running curl | bash!". That puts it into perspective what does security and privacy really mean? For all I know, these CLIs could have been secretly scraping my entire directory. Who's checking?
I only use agentic development tools in qemu VM that has no production
access or host shared folders, and the only secrets there are development.
I still have my Git ssh key there, but I never use ssh agent to load the keys.
The worst that agentic tools can do to me, is to install keylogger and compromise
my ssh key and password, but even then my deployment workflow includes creating release in Github,
which is not possible via ssh key only. I was very lucky to avoid the hack yesterday, as I use Claude installed via OpenVSX, but even if I installed the compromised version, the damage would be minimal.
I am doing similar with Debian 13 images in QEMU. Only access is to code specifically SSH in. On some, I am taking it a step further and my SSH key is not present at all on the bot. I add it as a git remote from my host into the VM and push and pull from there and then push to GitHub myself. So the agent has no path to GitHub either.
Wow, great notion about using git remote to transfer source from the agent host. I already use ssh into VM to transfer the files like designs for sharing, so going to adopt your approach right now :)
Then I just SSH in or use the vm gui and interact with the agent there. I like to keep it running in byobu so I can remote in more easily when I am on the laptop out and about.
Hell yeah that's the most sane and efficient personal dev sandboxing work flow I've heard and shows you know your stuff. Just basics but executed very well.
Don't forget: To make LLM useful you give it EVERYTHING it needs to know to be useful and if you don't use a certain provider with proper legal status, this is just gone.
We probably transfered more code/codebases and company secrets to random endpoints than ever before.
This was a compromise of the extension. The same kind of supply-chain attack could happen to literally any extension, package or other piece of software you install. Not excusing Anthropic letting this happen, but it doesn't have any bearing on the security and privacy of using Claude Code or Codex. Yes, you need to be careful about what they do. The fact that you might still get pwned via some other vector is irrelevant.
Can't their millions of agent swarms detect a malicious supply chain vector, sure it can happen to any company but not the one people are literally handing over their entire companies too.
Anthropic holds the SLA/MSA agreements not this bad actors what happens to your data is unknown if these become frequent and accepted. Given how much data is piped to Claude code via these services/extensions.
Both is completely secure theoretically. You can read the output of the curl before you refetch and pipe it to bash. You can also look at the commands claude is executing.
I think the difference people were warned about is that the server can infer the client, so there would be a difference between reading it in your regular browser and curling. As the other commenter says, I'm not sure how the server would be able to tell what curl does with the response.
Development must occur in an isolated VM, with no access to the publishing workflow. CI operations must happen in a separate VM or on the host system after scanning all files. Thus, even if the development tools are compromised (like this case), the impact is limited.
Still, it’s hardly a surprise that best practices are rarely followed with any consistency.
P(doom) crowd say Anthropic has not zero chance to destroy humanity.
What it is, the VS Code Extension Marketplace, including Cursor's, must be bigger.
Reading the report alone:
"First, the payload makes sure ~/.ssh exists and appends an attacker-controlled Ed25519 public key to ~/.ssh/authorized_keys, setting permissions to 0700 on the directory and 0600 on the file."
A system where ~/.ssh/authorized_keys has not been modified might be a hint that this did not execute - the article does not mention self cleanup/reverting to previous status behavior by the malware. By default /var/log/auth.log is only of marginal value because outbound SSH connections are not logged by default.
> Claude Code for VS Code extension on the Cursor Marketplace
So Cursor is running its own extension repository? Or is this the official VS Code repository and the incident report just happens to focus on Cursor for some reason?
VS Code doesn't allow forks to use their marketplace, I think. At least with VSCodium, they don't. I think they even patch a workaround where you could download almost all extensions' .vsix files.
1. If only Cursor Marketplace version was compromised, how was it published in the first place?
2. If the Cursor Marketplace version was released from an "official Anthropic" account, whose to say OpenVSX and/or VSCode versions are not compromised?
The affected versions have been pulled down, so no one really knows. The way I see it, only Cursor has admitted to have been compromised.
either the registry is compromised or someone at anthropic got compromised (or did it on purpose to steal crypto). the line in the cursor disclosure about the extension not being developed by anthropic is more confusing than clarifying. amateur hour for these massive corps
Something very similar is true for ChatGPT on macOS. I was recently very surprised to learn that starting a "Work" conversation on my Mac would grant the agent session file system access which can later be used remotely, e.g. from a browser session logged into the same account on a different computer.
Given how many people likely share ChatGPT subscriptions, this seems highly problematic.
We still don’t know exactly what happened, but I publish software to public registries, and supply chain attacks scare the hell out of me. They also frustrate me, because most of the defenses aren't automatic and often really hard to get right. We're really failing as an industry to do better here.
This is why I started to move more and more into my tix.im environment. You can't trust anything anymore - I assume everything is compromised if not proven otherwise.
It does. The compromised extension was published about 11 hours before Binance's coins went zoop, Binance blamed a compromised developer laptop, and OP pointed out that their hiring page explicitly notes developers will be using Cursor with Claude Code for "development efficiency".
Has anyone managed to confirm the time that the backdoored version was released? None of my machines updated to it, I am wondering if it was saved by a combination of luck and
> This extension is not updated yet because new versions become eligible for automatic update 12 hours after they are published. It will become eligible in 2 hours.
Yes, but they normally have a 12-hour delay (it was 2 hours at the beginning of June and got extended sometime after that, maybe later in June) and require some human action to reload extensions or restart the editor before the new version runs.
I noticed that there was no update yesterday and wondered why. The daily updates are annoying in the base case.
Edited: This update was live for only 11 hours, so hopefully the default delay limited the blast radius, as it is meant to do.
What I dont understand is why the hell do these extensions & clis need to update every damn day. clearly there's no new models everyday, what the actual f.
This is embarrassing to Anthropic given here that they lack basic security procedures to prevent this and they have gotten themselves hacked to push malware.
Claude Code was already malware, but with this attack you cannot trust the vendor either.
In the next 2 weeks, no one should download / install / update anything as I predict at least one developer's key for a GitHub, NPM, Crates.io or any other CI or login was leaked and the attacker will target those services to propagate their trojan to more victims.
It tells you all you need to know about how unserious these people are about computer security.
> I've recently talked to people at Anthropic who are very confident that they have effectively solved the prompt injection risks that cause lethal trifecta attacks with their latest models.
> The open challenge is how to prove it!
https://news.ycombinator.com/item?id=50297369
Or, is this FUD?
It also sounds like only cursor marketplace is affected, not vscode proper.
It may be the cause, but zero excuse.
How could Anthropic have known, they only claim to harbour superintelligence.
/s
It obviously doesn't take much to simply add 'just one more feature' that happens to undermine the original purpose of the application.
https://news.ycombinator.com/item?id=50302334 ???
"The Microsoft VSCode extension for Claude is NOT malicious"
"The Microsoft VSCode extension for Claude is NOT malicious"
https://opensourcemalware.com/blog/trojanized-claude-code-ex...
why are you spewing straight up lies?
"The Microsoft VSCode extension for Claude is NOT malicious"
dang should change the title and/or merge this submission with the other one. [1] This is causing needless panic for a typosquatting-style "attack" unrelated to official Anthropic extensions.
[1]: https://news.ycombinator.com/item?id=50301983
https://phorion.io/blog/offical-claude-code-vsx-compromise/
> Cursor said the extension was not developed by Anthropic and the maliciously modified extension was live on the marketplace from about 2:27 a.m. PT to 1:35 p.m. PT on Thursday
So is this or isn't this from Anthropic and how did this happen?
Big companies are not just 'Company' they are a lot of small teams doing their own things. They need to coordinate with other people for public announcements etc. and security team is probably also looking at it.
I think you’re giving Musk too much credit.
SpaceX will be re-valued on Anthropic's, or OpenAI's success, he, and his people are smart enough to understand this.
So really, there's no way to say what was compromised and where.
This could have been avoided by having extensions signed by company doing them with their own keys and have them reviewed
So that’s how these things can go, the employee’s machine may have been infected which was then used to inject this.
Looks good to me...
Nobody knows ... what any of it means... sing
[1] https://m.youtube.com/watch?v=3TNpOD6bov8
In all seriousness, there is nobody at Anthropic who can even give you that answer, because nobody reads their own code anymore. Take a look at the previous Claude Code leaked codebases, there is hundreds of files with 1000+ or more import statements at the top.
> Note: This delay does not apply to extensions from trusted publishers such as Microsoft, GitHub, and OpenAI. These extensions continue to update immediately.
You can't even configure which publishers you "trust"...
EDIT: I am now actually unsure if this is the case. From what I can see, all of the write-ups mention the Cursor marketplace and Open VSX specifically, but make no explicit mention of the official Microsoft registry.
EDIT 2: The version that was compromised (2.1.313) seems to have never been published on the official Microsoft registry. As per the documentation [1], unpublished versions should still be discoverable:
> When you unpublish an extension, the Marketplace preserves the extension statistics. The extension remains publicly discoverable and available via an existing API.
[1]: https://code.visualstudio.com/api/working-with-extensions/pu...
This very much illustrates the danger of just carelessly installing untrusted software on your machine. In this case it happens to be code from Anthropic, but it could have been almost anyone. I'll be very interested in see the explanation from Anthropic, because of all the extension developers they should have the best tools for finding issues like this. LLMs have proven pretty good at finding security issues, so did they simply not let their own AI tools vet the code before shipping it? Or was the extension infected later, because in that case every single extension on Open VSX could be at risk.
This would actually be an exceptionally good use-case for AI. You don't need a super-duper frontier model to download a bit of source-available code and check for suspicious signs, any cheap model would do well enough. Heck, it's effectively virus scanning, we've had that for decades.
https://github.com/jeanp413/open-remote-ssh
and run vscode-server inside a VM or container
I only have a single "Local Extension" and it is open-remote-ssh itself.
Of course anything you put into the container can be compromised, like git ssh keys or cloud provider keys/auth-tokens. But at least it keeps your PC from becoming a botnet and your browser credentials out of reach. You can easily and if you have a Containerfile with everything you need quickly wipe everything and re-create your dev environment (which I do every once in a while).
1. The Cursor Marketplace (never heard of this one, it’s possibly a custom marketplace hard-wired into Cursor, which is architecturally a VS Code fork?)
2. The Visual Studio marketplace, hard-wired into VS Code
3. Open VSX
If I understand this thread correctly, then #1 and #3 had compromised versions, and nobody has made a statement yet about the status of #2.
Edit: Cursor is apparently stating now that the tampered files were only on their mirror not OpenVSX.
https://cursor.com/blog/claude-code-extension-security-notic...
Edit: related https://news.ycombinator.com/item?id=50301983
Wtf does that mean.
``` host claudebox hostname 192.168.122.105 user developer IdentityFile ~/.ssh/id_rsa ```
``` host codexbox hostname 192.168.122.106 user developer IdentityFile ~/.ssh/id_rsa ```
I drop only my public key inside the VM. Then I set up the ~/Projects folder with the code the bot is working on any any authorized samples.
Then in my local on my host clone of the repository, I add another remote like
``` git remote add codexbox ssh://codexbox/home/developer/Projects/hello-world ```
Then git push codexbox.
Then I just SSH in or use the vm gui and interact with the agent there. I like to keep it running in byobu so I can remote in more easily when I am on the laptop out and about.
We probably transfered more code/codebases and company secrets to random endpoints than ever before.
Btw. you could setup a sandbox and limit access
Anthropic holds the SLA/MSA agreements not this bad actors what happens to your data is unknown if these become frequent and accepted. Given how much data is piped to Claude code via these services/extensions.
Yeah I sympathize with people saying software developers should not be called engineers too.
I just think that the level of malpractice in software engineering is a lot higher than in other engineering fields, for some reasons.
You always do that, right? ...right?
So you could be diligent and curl into a text file, see nothing is wrong, but then when you curl into bash, it runs something different.
An official announcement from Anthropic would be nice!
Still, it’s hardly a surprise that best practices are rarely followed with any consistency.
A system where ~/.ssh/authorized_keys has not been modified might be a hint that this did not execute - the article does not mention self cleanup/reverting to previous status behavior by the malware. By default /var/log/auth.log is only of marginal value because outbound SSH connections are not logged by default.
So Cursor is running its own extension repository? Or is this the official VS Code repository and the incident report just happens to focus on Cursor for some reason?
2. If the Cursor Marketplace version was released from an "official Anthropic" account, whose to say OpenVSX and/or VSCode versions are not compromised?
The affected versions have been pulled down, so no one really knows. The way I see it, only Cursor has admitted to have been compromised.
Given how many people likely share ChatGPT subscriptions, this seems highly problematic.
> This extension is not updated yet because new versions become eligible for automatic update 12 hours after they are published. It will become eligible in 2 hours.
Notable customers who run Claude Code and Cursor on laptops include Binance, amongst thousands of others.
Did the Claude Code plugin hack cause the Binance theft? The timing seems like an incredible coincidence.
You can try not installing malware like Claude Code or, if you absolutely must, doing it in a sandbox.
I noticed that there was no update yesterday and wondered why. The daily updates are annoying in the base case.
Edited: This update was live for only 11 hours, so hopefully the default delay limited the blast radius, as it is meant to do.
Claude Code was already malware, but with this attack you cannot trust the vendor either.
In the next 2 weeks, no one should download / install / update anything as I predict at least one developer's key for a GitHub, NPM, Crates.io or any other CI or login was leaked and the attacker will target those services to propagate their trojan to more victims.
Hats off to the hackers.