Cursor Traces Its Poisoned-Extension Breach to a Vendor Key and Old Agent Traces Holding Live Credentials

In a November 13, 2026 incident update, Cursor said an attacker used an access key exposed by Fireworks to copy historical coding-agent training traces, harvested unredacted credentials from them, and later used a stolen credential to plant tampered Claude Code extension builds in its marketplace cache on October 29.

A bronze key rests on yellowed files in an open archive drawer as a thin beam of light falls through a half-open door in a moss and smoke-blue records room.
A bronze key rests on yellowed files in an open archive drawer as a thin beam of light falls through a half-open door in a moss and smoke-blue records room.

Cursor has published the root cause of the breach that put a tampered Claude Code extension in its marketplace. In an update dated November 13, 2026, the company said it did not start with a flaw in Cursor’s editor or in Anthropic’s software. It started with an access key Cursor had issued to Fireworks, a third-party training and inference vendor. Cursor says Fireworks exposed that key, and an attacker used it to copy data from a storage bucket the two companies shared.

The bucket held historical coding-agent traces, which then gave the attacker the credentials for the rest of the intrusion.

From a shared bucket to the marketplace

Cursor’s account runs as a chain:

  1. The vendor key. The attacker used the Fireworks-held key to copy data from a read-only bucket Cursor used to share data with the vendor. It contained a limited set of coding-agent traces, mostly from 2025, from sessions not using Privacy Mode, which Cursor was permitted to use for model training.
  2. Credentials inside the traces. A small number of those traces contained credentials that had not been redacted. Cursor says less than 0.01% of customers had API credentials exposed this way.
  3. Lateral movement into SaaS tools. Using those credentials, the attacker reached select SaaS tools Cursor uses. From there they accessed some archived, non-production Cursor source 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, and some credentials users had sent to Cursor’s support team.
  4. The poisoned extension. On October 29, with a credential stolen from that exposed internal data, the attacker placed malicious copies of Anthropic’s Claude Code extension directly on the cache that serves Cursor’s VS Code extension marketplace. The upstream source, Open VSX, never carried the malware, and Anthropic did not publish the copies.

Cursor says at most 5,000 users who requested specific cached versions on October 29 may have received the malicious build, which was live for about 11 hours. The tampered versions were Anthropic.claude-code 2.1.312 and 2.1.313. Cursor warns that uninstalling the extension alone may not be enough and points affected users to the remediation steps in its October 30 security notice.

The company says the last attacker activity it saw was on November 1, and that attempt was blocked. It believes the incident is contained.

What Cursor says stayed out of reach

The update lists systems it says were not affected:

  • prompts and code from users with Privacy Mode enabled;
  • live traffic and active inference requests, since the exposed traces sat in storage for historical offline analysis;
  • production systems and infrastructure;
  • the Cursor app and its official updates, which ship from a separate code-hosting system the attacker never reached; no release-signing keys were exposed;
  • every other extension in the marketplace, which Cursor says it validated before rebuilding the store, and plugins from the Cursor Marketplace;
  • customer code repositories.

For users not on Privacy Mode, Cursor says coding-agent traces were exposed for fewer than 0.05% of users, and that each has been notified directly. Team admins whose credentials leaked were emailed lists of specific keys to rotate, and Cursor asked the services that issued those keys to revoke them.

Moving inference off Fireworks and rotating every key

Cursor’s remediation list starts with the vendor relationship. It moved all inference traffic off Fireworks, deactivated the key and cut the vendor’s access to every shared dataset. It rotated every exposed credential it found, highest risk first, and then rotated all credentials in use across internal systems as a precaution. Other steps include tighter network restrictions on service providers so stolen credentials cannot be used outside their intended location, secret scanning across all code repositories and other internal data sources, a precautionary redeploy of production services, and outside forensics support.

The incident shows a risk that grows as coding agents do more work. Agent traces record what an agent saw and did, and that can include tokens, keys and passwords pasted into a terminal or read from a config file. When those traces are kept for training and shared with an outside vendor, a leaked vendor key turns into a credential dump. Cursor did not say how Fireworks exposed the key, and the update does not include a statement from Fireworks.

Topics
  • #Cursor
  • #Supply Chain
  • #Incident Response
  • #AI Coding Agents
  • #Credentials
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.