TL;DR
Cursor says a key exposed by its AI vendor let an attacker copy old coding-agent records. Credentials in those records led to other Cursor systems.
The attacker later used a different stolen credential to alter Claude Code extension copies in Cursor’s download cache. Up to 5,000 users potentially received them.
I’d check retained AI data for secrets, follow what those secrets can access and verify what reaches developers’ machines.

A key issued to an AI vendor ended up being part of an attack on a software extension marketplace.
That’s the chain Cursor described on November 13, after investigating an incident that first came to light on October 29. The attacker didn’t need to compromise Cursor’s official app releases or Anthropic’s extension publisher. Cursor says the attacker used access to data from a vendor relationship, found credentials in that data and eventually put malicious copies of an extension into a cache serving downloads to users.
The distance between those two systems is what makes this incident worth examining. How did access granted for one purpose reach software being downloaded by customers?
Start with the key and the old traces
In Cursor’s November 13 incident update, the company says an attacker obtained an access key Cursor had issued to Fireworks, an AI provider. Cursor says Fireworks exposed the key. The attacker used it to copy data from a read-only storage bucket used for sharing material with that provider.
The bucket held a limited set of coding-agent traces: records of interactions with a coding agent, mostly from 2025. Cursor says those sessions weren’t in Privacy Mode and it was permitted to use the material for model training. These were stored records, not live requests to AI models.
Some of the traces contained credentials that hadn’t been removed. Cursor says fewer than 0.01% of customers had API credentials exposed. It separately reports that coding-agent traces were exposed for fewer than 0.05% of users not in Privacy Mode. Not every Cursor customer had code or secrets exposed.
The attacker used credentials found in the data to reach selected online services Cursor used internally. According to the company, that access included limited support records, email conversations, product usage details, summaries of some coding-agent conversations and archived non-production source repositories. Some support records also contained credentials users had shared while seeking help.
This is where I’d examine the vendor arrangement: what the provider needed to read, why the data was retained and how credentials ended up in it. I’d also check whether a stolen credential could be used from anywhere on the internet.
I’m not saying that keeping training data automatically causes a breach. Cursor says its data sharing was permitted. The security question is whether the material was safe to retain and whether a stolen key could reach more than the vendor needed.
Fireworks has published its own November 9 incident update. It describes unauthorized activity involving internal systems and some customer data, with an investigation still underway. That account helps explain the vendor’s broader response, but it doesn’t independently reconstruct every step of the Cursor attack. I’m using Cursor’s investigation for the specific chain it reports.
Then the attack reached a download
On October 29, Cursor says the attacker used a different stolen credential, obtained from its internal systems, to place unauthorized copies of Anthropic’s Claude Code extension in Cursor’s extension marketplace cache, where files are stored for download.
Cursor says the malicious copies were not published by Anthropic and were not present in Open VSX, the marketplace’s upstream source. Cursor’s official app and its release-signing keys weren’t compromised either.
A customer could still request an extension through Cursor’s marketplace and receive a malicious cached copy of something they expected to be legitimate.
Cursor identified two affected versions, 2.1.312 and 2.1.313, on macOS and Windows. Its October 29 security notice says the modified extension tried to send device information to the attacker, add a key for SSH remote login and leave a way to reconnect later. Cursor says attempts to enable SSH failed on most macOS and Windows hosts because the extension lacked administrator privileges. The company says the attacker-operated server is no longer functional.
Microsoft Security Research independently examined the injected extension code and confirmed the attempted SSH key change, remote tunnel and persistence behavior.
The altered versions were available for about 11 hours. Cursor estimates that at most 5,000 users requesting particular cached versions could have received them. Cursor says it notified affected customers directly.
I’ve written about the need to know your suppliers and their access paths in You Can’t Patch a Supply Chain You Can’t Inventory. This incident adds an uncomfortable detail: even knowing which package you intended to download doesn’t prove the cache handed you what you expected.
What the original permission could reach
In Your AI Workflow Is on the Internet. Who’s Responsible for It?, I asked what a compromised automation service could reach through its connected accounts.
Here, the first key only allowed reading. The credentials in the retained data led to other systems, and a different stolen credential later allowed the attacker to change extension files. A review of the first key’s permissions wouldn’t have shown that whole route.
For a security team adopting an AI coding service, I’d ask the vendor owner and the engineering team to show three things.
First, what leaves our environment and how long it stays. Privacy settings and training permissions aren’t footnotes. Determine which prompts, code, agent traces and support submissions can be stored or used by a provider. Check whether the retained records can contain access tokens, passwords or personal data. Secret scanning should apply to traces and support attachments, not only to Git repositories.
Next, what a stolen vendor credential can actually read. Read-only access isn’t harmless if the data contains other working credentials. Limit each key to the data the vendor needs, restrict where it can be used when feasible and make expiration and rotation part of the arrangement. Look for credentials that have outlived the original session or project. If one turns up in a retained trace, have its owner revoke or rotate it at the service that issued it, then investigate where it was used.
Finally, what customers receive from a distribution system. The download cache needed its own access controls and integrity checks. Teams should examine how cached files are verified against the publisher’s authenticated originals, who can modify them and how customers can independently verify what they’ve downloaded through cryptographic hashes, digital signatures and trusted provenance. A matching file extension and version number don’t prove it’s the right file.
Those aren’t controls that can all be delegated to the AI vendor. The provider is responsible for its own environment, the developer-tool company owns its data-sharing and distribution systems and customers still need to know which extensions and versions are permitted on their machines.
I’d also ask whether the organization has a practical way to identify affected installations and revoke exposed credentials without relying on developers to notice a blog post.
What this incident does and doesn’t tell us
Cursor says the vast majority of its customers weren’t affected. Prompts and code submitted in Privacy Mode weren’t part of the exposed trace data. The company also reports that its official app releases, customer code repositories and live inference requests weren’t compromised. It removed the malicious extension, rebuilt its marketplace and rotated exposed credentials, along with other internal credentials as a precaution.
Cursor says the last attacker activity it observed was on November 1, and that attempt was blocked. That’s its finding as of November 13. It doesn’t independently establish that every affected system is clean.
If you installed versions 2.1.312 or 2.1.313 of the Claude Code extension from Cursor’s marketplace on October 29, follow Cursor’s remediation instructions and review Microsoft’s investigation. Don’t assume a machine is clean just because the suspicious SSH key isn’t there. Uninstalling the extension may leave malicious startup scripts and other persistence mechanisms behind.
I wouldn’t use this incident to argue that companies should stop using AI coding assistants or outside model providers. That’s not what the evidence shows.
What happened was more specific. Credentials that should have been redacted remained in stored data. An attacker used exposed credentials to move beyond the first compromised service, eventually reaching a system customers trusted to deliver software.
What data does the vendor receive? Where are copies stored? Who can retrieve them? And if someone steals a credential exposed through that relationship, what else can they reach?
If the security review ends at the vendor’s boundary, we’re missing part of the risk.
Resources
Cursor, An update on our recent security incident (November 13, 2026)
Cursor, Security notice for malicious Claude Code extension versions (October 29, 2026)
Microsoft Security Research, Inside a malicious Claude Code-branded extension (October 30, 2026)
Fireworks, Update on October 2026 security incident (updated November 9, 2026)

