Cursor says the malicious Claude Code extension that appeared in its marketplace did not start with Anthropic. It started with an access key Cursor had issued to Fireworks AI, an AI infrastructure provider, and it reached Cursor's marketplace only after the attacker used data from a shared storage bucket. Anthropic did not publish the tampered copies, and Cursor says the product itself was not vulnerable.
What happened, and when
Cursor first disclosed the incident on October 29, 2026, and published a fuller investigation on November 13. According to Cursor's security update, an attacker obtained a key Cursor had issued to Fireworks, the key was exposed, and the attacker used it to copy data from a read-only bucket that Cursor shared with the vendor.
The second stage came on October 29. Cursor says the attacker used a credential taken from that exposed data to place unauthorized copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's extension marketplace. BleepingComputer's coverage reports that Cursor has since revoked and rotated the exposed credentials, moved all inference traffic away from Fireworks, rebuilt its extension marketplace and redeployed its production infrastructure as a precaution.
A key handed to one party can still end up in someone else's hands. For the wider problem of credentials and agents, see our piece on rotating AI API keys after a breach.Why the vendor link matters more than the extension
The affected builds were versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension. Cursor's update tells customers who installed or updated those versions from its marketplace during the exposure window to follow the steps in its security notice. If you were on Cursor at that time, do not assume the problem is someone else's to solve.
The headline makes this sound like a Claude Code problem. It is really a supply-chain problem with three weak points: an outside vendor that exposed a key, a shared data bucket that held more than it should have, and a marketplace cache that served a build nobody at Anthropic had published. Our earlier pieces on LiteLLM shipping a credential stealer and on prompt injection as a supply-chain problem describe the same pattern from different directions.
What it means for solo operators
You are probably not operating a code editor marketplace. You are probably operating a stack of AI tools, each holding a credential you set up once and forgot. Three moves matter this week:
- Inventory every vendor key. List each service that holds a credential of yours, including the inference providers you only use through someone else's tool. Rotate on a calendar, not only after an incident.
- Check where your extensions came from. Confirm the publisher, version and install source of every editor extension, especially any that talk to AI services. An extension with the right name is not the same as the right build.
- Act on the version numbers in the notice. If an incident report names specific builds, check whether you installed them, and follow the steps the vendor gives. A quick uninstall is a guess, not a remediation.
Our solo operator security checklist for an agent stack turns that inventory into a routine you can run in an afternoon.
Our take
After more than 25 years in digital marketing, we have learned that the most expensive vendor problems rarely arrive as a dramatic hack. They arrive as a forgotten login, a shared folder nobody audited, or a plugin everyone clicked install on because it looked official. This incident has all three flavors, which is impressive in a way nobody asked for.
Let us be clear about the boundaries. Cursor's own account puts the root cause at a key Fireworks exposed and a shared bucket that held credentials it should not have held. Anthropic's product, by Cursor's account, was not the vulnerability. Blaming the model maker here would be lazy, and so would shrugging it off, because real developers may have installed tampered code and real credentials were exposed. Taking the problem seriously is not optional.
Our advice is boring on purpose: fewer keys, shorter lifespans, fewer extensions, and a habit of reading the incident notice before you reach for the reassurance. The industry keeps promising that agents will manage our security for us. For now, the agent is only as trustworthy as the least careful vendor in the chain, and that vendor is usually one you have forgotten you hired.