SI Agent News
DevelopingClaudeSecurity

Stolen Fireworks AI key leads to Cursor data breach and trojanized extension

A security lapse at Fireworks AI allowed attackers to steal Cursor data and distribute a malicious version of Anthropic's Claude Code extension.

Documentary photo: picture for the story 'Stolen Fireworks AI key leads to Cursor data breach and trojanized ext'
AI-made documentary photo for this story. No real people or products are shown.
0.01%percentage of customers with exposed credentials

The short version

Why it matters

This highlights the risks of third-party supply chain vulnerabilities where a breach at one partner can cascade into multiple downstream services. It serves as a reminder that even 'read-only' data shares can contain sensitive credentials if not properly scrubbed.

The Fireworks AI Breach

The security incident began when an attacker obtained an access key issued to Fireworks AI. This key was exposed during a breach targeting Fireworks, which allowed the attacker to access a storage bucket used by Cursor. Cursor used this specific bucket to share data with the AI company.

The data in the bucket consisted of coding agent traces from sessions that were not protected by Privacy Mode. Privacy Mode is a setting in Cursor that allows users to opt out of having their data used for model training. Because these sessions were permitted for training, they contained some information that was not intended for public view.

Among the exposed traces were credentials that had not been removed from the logs. These credentials exposed the API keys of less than 0.01% of Cursor's customers. The attacker used these starting points to reach various software-as-a-service tools and internal systems used by the platform.

Data and Source Code Exposure

Once inside the system, the attacker accessed several types of sensitive information. This included archived non-production source code repositories, some of which contained business contact information. The breach also reached limited customer support data and various email conversations.

Other stolen items included product usage metrics and specific contact details. The attacker also managed to gather summaries of some coding agent conversations and credentials that users had sent to the Cursor support team. These findings indicate a significant reach into both internal operations and customer interactions.

Cursor responded by deactivating the vendor's key and rotating the compromised credentials. They also began a hunt for signs of deeper access to contain the threat. To protect the platform, the company temporarily disabled internal and product services, which disrupted some features for active users.

The Trojanized Extension

The attack escalated on October 29, 2026, when the intruder used a stolen internal credential to target the extension marketplace. The attacker placed a trojanized version of Anthropic's official Claude Code extension directly onto the cache used to serve the marketplace. A trojan is a malicious program that appears to be legitimate software.

This malicious build attempted to install a persistent SSH backdoor on the systems of users who downloaded it. An SSH backdoor allows an attacker to maintain remote access to a computer at any time. The trojan was available for approximately 11 hours before the company identified and removed it.

Cursor reported that roughly 5,000 people who pulled specific cached versions of the extension on that date might have gotten the bad build. The company took the entire extension marketplace offline to remove the threat. They also worked with the cloud service provider hosting the attacker's server to shut it down.

Impact and Remediation

Cursor notified the users who installed versions 2.1.312 or 2.1.313 during the exposure window. The company provided assistance to those users to help them revoke any credentials that might have been exposed. They also moved all inference traffic away from Fireworks and cut off all shared data sets.

The company stated that they have not seen any further signs of unauthorized activity since the start of November 2026. They also noted that there is no proof that people who used Privacy Mode were affected by the breach. Furthermore, the production environment, the Cursor app, and other plugins downloaded from the marketplace remained unaffected.

This incident is part of a larger pattern involving Fireworks AI. Uber confirmed that a service provider—identified by reports as Fireworks—lost internal credentials, leading to the theft of thousands of drivers' license records. Additionally, the productivity startup Notion confirmed its data was also stolen during the Fireworks AI hack.

What to watch

Try it

Check your Cursor extension versions and ensure you are using the latest updates to avoid cached malicious builds.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by The Hacker News · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news