1 party, 2 party, 3 party, 4
Cursor, a popular code editor and agentic AI tool owned by SpaceX, has just published a report about its recent report of a data breach.
As the company puts it:
On October 29 [of 2026], we shared that we were investigating a security incident and a malicious, unauthorized copy of Anthropic’s Claude Code extension that appeared in the Cursor VS Code extension marketplace.
After an extensive investigation over the past two weeks, we want to share more context on the attack and the related impact.
As seems to happen so often these days, the Cursor breach was caused by a by a breach at a third party called Fireworks, who in turn blamed the breach on a fourth party that it used as a cloud hosting provider.
As the old childrens’ counting song puts it, “One potato, two potato, three potato, four/Five potato, six potato, seven potato, more.”
For affected Cursor customers, of whom there were about 5000, the attack ultimately led to them installing a booby-trapped but apparently official version of an Anthropic plugin from Cursor’s own online marketplace.
Cursor’s software is based on the super-popular Microsoft Visual Studio Code editor (VS Code), which is in turn based on the Electron software framework, which is based on a combination of the Node.js JavaScript engine (intended for full-blown software that runs outside a browser), plus the Chromium web browser for its user interface. Cursor provides its own supposedly trustworthy source of VS Code extensions, which is why customers were inclined to trust the malware-laced Anthropic extension that showed up there.
Bypassing the firewall
The malware implanted by the bogus extension had two primary aims:
- To steal data about you and your computer, including “hostname, username, operating system, architecture, home directory path, environment variable names, working directories, configured Git email, and public IP [number].”
- To set up an SSH login on your computer, protected by a cryptographic key created by the crooks themselves, so they could sneak back in later.
The first side-effect is bad enough, not least because so-called environment variables, which form a database of useful values kept in memory rather than saved permanently to disk, may identify the online services you use, and reference authentication tokens or other cryptographic secrets.
The second side-effect is worse, because the infection process sets up what’s known as a reverse tunnel back to a server operated by the attackers.
Their aim is to get long-term remote access to your computer, while avoiding any firewalls or security appliances in the way.
Simply put, a reverse tunnel is a network connection set up back-to-front, so it only ever connects outwards (which is usually allowed), instead of passively waiting for a connection to come inwards (which is typically blocked).
Setting up a network connection generally involves a computer that kicks things off, typically called the client, and a server, which is listening out for and accepting those incoming connection attempts. (When you surf the web, the client is your browser and the server is the website you’re visiting.)
But once the connection has been set up, it’s bidirectional, and each end effectively becomes an equal partner in the link, known in the jargon by the term peer-to-peer, which reflects the surprisingly egalitarian nature of most low-level connections on the internet.
Everything shoved into the network connection at one end emerges byte-for-byte at the other, and vice versa, so either end can behave as a the client or the server in whatever happens next.
A reverse tunnel, as mentioned above, is a client-server interaction that’s established via an outbound connection from the computer that will ultimately behave as the server, and accepted as an inbound connection by the computer that will ultimately behave as the client.
In this case, the victim’s computer ends up as an SSH server that the criminals intend to log into from their end once the connection is alive and running “in reverse,” thereby hoping to sidestep any firewalls that would otherwise be in the way.
How bad was it?
Fortunately, for many victims, Cursor found that the part of the malware that opens up and configures a new SSH login didn’t work, due to sloppy coding on the part of the malware creators.
Sometimes, software bugs work out for the best.
That wasn’t all, however.
Because this attack involved the criminals breaking into the Fireworks network first, those criminals were able to snoop around in customer data held on the breached server, which included data that Cursor and its customers had shared with Fireworks.
Customers who hadn’t turned on Cursor’s “Privacy Mode” feature had apparently shared data for Cursor to use in training its AI agents, and some of that data included passwords and access credentials.
Those who had turned “Privacy Mode” on were less likely to have been affected, but Cursor admitted that some of them had their access credentials harvested by the criminals anyway.
Of course, as we know from the subsequent appearance of the booby-trapped extension on Cursor’s own servers, some of the credentials stolen in this stage of the attack were Cursor’s own.
As well as getting into Cursor’s extension publishing system, it seems that the attackers also got into some of Cursor’s customer databases and made off with “business contact information, limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding agent conversations, and some credentials users had sent to the Cursor support team.”
What to do?
- If you’re a Cursor customer, you should have been told if you were affected, either because you downloaded the malicious extension onto your own computer, or because your data was accessed via Cursor’s cloud. But Cursor may not be able to tell definitively who was breached, so please read the company’s reports for yourself.
- If you’re worried about the malware because you use the Anthropic extension, check out Cursor’s advice on how to check whether you were affected. (They have created a script to clean up the malware if you were.)
- If you run AI agents that might share data for the commercial benefit of your vendor, make sure that you have turned sharing off as fully as you can (or choose a service that never does it), unless you’ve made an informed, intentional decision to enable it.
- If you share data with third party service providers yourself, whether via the cloud or not, be aware before you share, and take care never to include password, access tokens or other authentication secrets by mistake.
Ask how SolCyber can help you do cybersecurity in the most human-friendly way. Don’t get stuck behind an ever-expanding convoy of security tools that leave you at the whim of policies and procedures that are dictated by the tools, even though they don’t suit you or your business!
More About Duck
Paul Ducklin is a respected expert with more than 30 years of experience as a programmer, reverser, researcher and educator in the cybersecurity industry. Duck, as he is known, is also a globally respected writer, presenter and podcaster with an unmatched knack for explaining even the most complex technical issues in plain English. Read, learn, enjoy!