Cursor breach started at AI vendor Fireworks: leaked key exposed customer data and source code
A key Cursor issued to Fireworks leaked. The attacker copied training traces with unredacted credentials and reached support data, archived code and the fake Claude Code extension.


The attack on Cursor did not start at Cursor, but at a vendor. In an update published on 13 November, the company says an attacker obtained an access key Cursor had issued to Fireworks, its third-party training and inference vendor. With that key the attacker copied data from a storage bucket Cursor shared with Fireworks, and from there worked further into several of Cursor's own tools. It is the same incident that ended with a fake Claude Code extension in Cursor's marketplace on 29 October.
From one vendor key to customer data
According to Cursor, the key was exposed by Fireworks. Cursor does not say how. The bucket the attacker copied from was read-only and held a limited set of coding agent traces, mostly from 2025. The traces came from sessions where users had not turned on Privacy Mode, and which Cursor was permitted to use for model training.
Some of these traces contained credentials that had not been redacted. Cursor says fewer than 0.01 percent of customers had API credentials exposed this way. Using those credentials, the attacker reached select SaaS tools Cursor uses internally. From there the attacker accessed:
- select archived non-production source code repositories, some containing business contact information
- limited customer support data and email conversations
- product usage metrics and contact details
- summaries of some coding agent conversations
- some credentials users had sent to Cursor's support team
For users not on Privacy Mode, Cursor says coding agent traces were exposed for fewer than 0.05 percent of users. The company says all of them have been notified directly, and that anyone who has not received a notice was not in the traces.
How the fake extension reached the marketplace
Using a stolen credential from the exposed data, the attacker on 29 October placed malicious copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's VS Code extension marketplace. Cursor stresses that Anthropic did not publish the copy and that this was not a vulnerability in Anthropic's product. The upstream source, Open VSX, never had the malware.
Up to 5,000 users who requested the affected versions that day may have received the malicious build. It was available for about eleven hours. hogby.ai covered the extension itself on 31 October. What is new in the update is how the attacker got in, and which data was affected.
Cursor says the last attacker activity it has seen was on 1 November, and that it was blocked. The company believes the incident is contained but keeps monitoring its systems.
What was not affected
Cursor lists what it says was not affected: prompts and code for users with Privacy Mode, data being sent to Cursor now, production systems, the Cursor app and its official updates, customer code repositories, other extensions in the marketplace and active inference requests. Cursor says the app ships from a separate code hosting system the attacker never accessed, and that no release signing keys were exposed.
In response, Cursor has moved all inference off Fireworks, deactivated the key and cut off the vendor's access to all shared data sets. It has rotated every exposed credential and, as a precaution, all credentials across internal systems. Cursor has also added secret scanning to all code repositories, tightened network access restrictions at service providers, redeployed production infrastructure and brought in external forensics experts.
What leaders should take away
The incident shows how far a leak can travel when data is shared with an AI vendor. Data shared for model training contained secrets, and those secrets opened new doors. Three questions are worth putting to your own AI coding tool vendors: Are our sessions used for training, and by whom? Are keys and passwords removed before data is shared? And can a leaked key be used outside the place it is meant for?
For companies using Cursor, the concrete step is to check whether you have received a notice from Cursor, and whether Privacy Mode is on for your team. If admins received a list of keys to rotate, that comes first. Staff should also avoid sending passwords or keys to any vendor's support team.
Cursor does not give the number of affected customers in absolute terms, how Fireworks exposed the key, or when the intrusion began. Fireworks has not commented in the sources hogby.ai has reviewed.
Sources and media
- Cursor: An update on our recent security incident, dated 13 November 2026. Checked 14 November 2026.
- Cursor on X about the update, posted 14 November 2026 at 02:13 Oslo time.
- Cursor: Security notice – malicious version of Anthropic.claude-code (2.1.312 and 2.1.313), dated 29 October 2026.
Illustration: AI-generated with Grok Bot / hogby.ai