Anthropic got compromised.
-
@GossiTheDog here's a detection #osquery for #macos, in case it helps anyone
https://gist.github.com/RichardoC/9753dfa5aeffc6a1201f5e57114f54ff
@RichardoC - how do we run this? can you give a brief?
-
@GossiTheDog I expected a "Malicious official Claude Code extension on VScode" to be... louder?
I'm not sure if it was an incident beyond cursor marketplace. Enough time has passed to see more information if it was that large scale.@moh53n @GossiTheDog to be honest I expected VScode incidents to be more common, has that happy extension wild west feel like browser extensions before people started trading in them.
-
I've written a blog - Anthropic have been hacked to deliver remote access malware to millions of endpoints where developers use Claude Code.
@GossiTheDog This sentence needs a bit of proofreading:
"Anthropic has not informed customers, and the malicious version was silently deleted the backdoored version without telling anybody."
Also, it might be a good idea to include instructions what the potential victims show do, in order to detect whether they have been compromised:
macOS:
grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted"
Windows:
$files = "$HOME\.ssh\authorized_keys", "C:\ProgramData\ssh\administrators_authorized_keys" | Where-Object { Test-Path $_ }
if ($files -and (Select-String svc-telemetry $files)) { "svc-telemetry key found" } else { "No key found; machine not impacted" } -
@GossiTheDog What, Teams is down?
-
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog of course Anthropic got compromised.
In what reality was that ever not going to happen?
-
I've confirmed Binance use Claude Code on Cursor with this extension.
@molly0xfff
This thread might be interesting to you! -
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog Exactly why I wrote AI Playground. https://gitlab.com/cryptomilk/ai-playground
-
@GossiTheDog with no comms from Anthropic, are there any concrete recommendations? I can think of at least one person I know who might be impacted by this.
@huronbikes @GossiTheDog I'd go with:
Disconnect from any network
Save your data (only) to removable storage
Erase and reinstall from media.Essentially, 'Nuke it from orbit'
I recognise this might be a bit draconian for many, though. -
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog
Well I don't mind. -
FWIW the compromised version, 2.1.313, was removed from repos. 2.1.314 is the ‘fixed’ version.
@GossiTheDog only 2.1.313 was comprised, no previous versions?
-
I've written a blog - Anthropic have been hacked to deliver remote access malware to millions of endpoints where developers use Claude Code.
@GossiTheDog richly ironic.
-
I've written a blog - Anthropic have been hacked to deliver remote access malware to millions of endpoints where developers use Claude Code.
@GossiTheDog "Outside of highly unusual circumstances where Cursor was started with administrator privileges" wait don't the slop kiddies do
>sudo curl ... | bash
or
>docker import <url>
all the time? -
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
-
Cursor statement "This malicious version of the extension was not developed by Anthropic" is a bit wordsmithy. "Developed" is doing some heavy lifting - it was uploaded to Anthropic's verified account and Cursor automatically updates extensions.
Cursor say around 5k developer systems were impacted.
@GossiTheDog I don't find any indications that it was anything but a compromise of Cursor systems (proprietary plugin ingestion it seems) , I don't find anything that the official extension or OpenVSX would have carried the malicious version. What evidence have you found that it's bigger?
In addition VS Code has an default extension update block for 12h, so it wouldn't have been affected. -
RE: https://infosec.exchange/@RichardoC/117529759531517074
MacOS OSQuery hunting queries:
@GossiTheDog That phorion blog post only talks about the Cursor Marketplace version which I guess is a bit of an obscure place to get extensions from. Were the official Microsoft marketplace and OpenVSX affected?
-
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog At a certain point, a critical vulnerability is indistinguishable from an alignment problem.
-
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog@cyberplace.social LETS FUCKING GOOOOOOOOOOOOOOOOOOOOOOOOOOO
-
Anthropic got compromised.
Their Claude Code extension on Visual Studio has malware in it, which creates a remote access Trojan to desktops.
@GossiTheDog I think the compromise is only on the Cursor marketplace, not VS Code's. So users of Cursor are affected, users of VS Code should be fine. The name of the extension is "Claude Code for VSCode". That's how I'm reading the report. The article states "Claude Code for VS Code on Cursor Marketplace" which is confusing at first glance, and it's frustrating that they're not making is any clearer than that.
-
@RichardoC - how do we run this? can you give a brief?
@saiskandadotin If your Mobile Device Manager(MDM) (say Swif or Jamf) or your endpoint detection and response (EDR) such as crowdstrike supports running osquery on your machines, you can use it through that.
What MDM/EDR do you have and I can try and find some equivalent docs,
For context, this was written for folks in a security team in a company large enough to have some sort of fleet level management - rather than individuals trying to check their own machines
Details for swif MDM - https://help.swif.ai/en/articles/14103730-run-commands-by-script-osquery-file-or-git-repo
-
@GossiTheDog "Live by the sword, die by the sword" but not in quite the way that Musk and other muppets have put forth.
In Q1 this year I was getting ready to evaluate Claude Code as a product, when the news dropped that Anthropic had been prompting their models to lie about being human in contributions to open source repositories, at the same time as the source code leak for Claude Code itself. This was the "Carter Burke from Aliens" moment for me regarding that company. That they haven't even published a disclosure for this, tells all.
World's smallest violin. This stuff was always an INFOSEC Three Mile Island waiting to happen. I continue to NOT use "AI" products of any kind in my work and am more productive as a direct result.
In fact to do so would be actively deleterious and a complete waste of time as the models are not trained for novel work (they simply don't innovate, they can't, they're stochastic token predictors) and Herb Sutter pointed this out in March at BeCPP Symposium to boot, before any of this happened.
@bms While I am 100% with you directionaly, and am not an LLM defender in any way, I detest the damn things, I think there is a sense in which recombining the things that have come before absolutely can come up with novel combinations which transformatively advance the boundaries of what is possible - this is, after all, what sexual reproduction does, and how evolution works. I'm not saying you're wrong, only that *this* argument isn't the one to use against LLMs.
