High

Cursor, Claude Code, Anthropic, Fireworks AI, supply chain attack, VS Code extension, AI coding agents, secrets, Notion, Uber

Cursor Claude Code Extension Hack Traced to Fireworks AI Key

The malicious Cursor Claude Code extension (2.1.312, 2.1.313) traces to a key stolen in the Fireworks AI breach. Who is affected and what to do.

The malicious Claude Code extension that Cursor's marketplace served on October 29 traces back to the Fireworks AI breach. Cursor says an attacker stole an access key Cursor had issued to Fireworks, used it to copy agent traces that contained unredacted credentials, worked through Cursor's SaaS tools, and finally placed trojanized versions 2.1.312 and 2.1.313 of the anthropic.claude-code extension in the marketplace cache.

Cursor published the root cause in a Friday update, reported by The Hacker News and SecurityWeek. Anthropic was not hacked. All of the malicious activity happened in Cursor's environment, and the tampered builds were never in the Open VSX source; they were served only to users who requested the cached versions during an 11-hour window. Cursor puts the number of exposed users at no more than 5,000. The tampered extension tried to install a persistent SSH backdoor.

The chain started outside Cursor. Fireworks AI, an inference platform, was hacked twice in October: first through internal credentials that reached stored secrets and billing data, then through a compromised third-party cloud provider that led to Fireworks development machines. A Cursor key issued to Fireworks was exposed along the way. The attacker used it to copy a read-only bucket of coding agent traces from sessions not in Privacy Mode, most dating to 2025 and approved for model training. Some traces still contained credentials, exposing API credentials for under 0.01% of Cursor customers.

What did the attacker access at Cursor?

From the SaaS tools those credentials opened, the attacker reached archived non-production source code repositories (some with business contact information), limited customer support data and emails, product usage metrics, summaries of some agent conversations, and credentials users had sent to Cursor support. Beyond the marketplace cache, Cursor says there is no evidence of impact on Privacy Mode users, customer code repositories, its production environment, the Cursor app and its official updates, or any other marketplace extension.

Date

Event

October 2026

Fireworks AI breached twice; Cursor key issued to Fireworks exposed

Before Oct 29

Attacker copies agent traces, uses leaked credentials to reach Cursor SaaS tools

October 29

Trojanized anthropic.claude-code 2.1.312 and 2.1.313 served from marketplace cache for about 11 hours

October 29

Cursor removes the builds, takes the marketplace offline, gets the attacker's server taken down

November 1

Last attacker activity Cursor has seen

November 13

Cursor publishes root cause; moves inference off Fireworks

Cursor was not the only Fireworks AI customer hit. Notion confirmed its data was stolen, and Uber disclosed that a service provider's compromised credentials were used to reach Uber data; The Record reports nearly 150,000 people were affected, and TechCrunch identified the provider as Fireworks AI.

Am I affected by the malicious Claude Code extension?

Editor extensions run with the developer's full user permissions, so a malicious one can reach anything that user can. You are affected only if you installed or updated to anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace on October 29. Cursor has notified those users directly. It warns that uninstalling the extension alone may not be enough, because the payload aimed to set up persistence. Follow Cursor's security notice, review SSH authorized keys and startup items on the affected machine, and rotate any credentials that were reachable from it, including cloud keys, Git tokens and LLM API keys.

Our read

Two lessons stand out. First, agent traces are a new place for secrets to leak: credentials pasted into a coding session ended up in training data shared with a vendor, then in an attacker's hands. Second, every key you issue to a supplier extends your attack surface to theirs. Scan for secrets in code and in anything your AI tools log (the OWASP Secrets Management Cheat Sheet covers rotation and scoping), keep vendor keys narrowly scoped and short-lived, and review which suppliers hold access to your data. The OWASP Top 10 for LLM applications lists supply chain and sensitive information disclosure among its core risks. We covered a related leak path in Claude Code's local settings file and broader AI coding agent sandbox escapes. Our guide to AI coding assistant security risks covers controls for teams using these tools, and ARGUS scans repositories for leaked secrets and vulnerable code on every pull request.

Reporting by The Hacker News and SecurityWeek; root cause per Cursor's November 13 update; Uber figure per The Record.

Writing your own code with AI? The same bug classes surface there too. Scan your code free with Argus →
Liked this briefing? Share it:

More briefings

Get the briefings first
Breaking security news, verified fast, with the one fact the headlines skip. No spam - unsubscribe anytime.
AI CODE SECURITY
Catch the bug before it ships
Argus scans your repos for the vulnerability classes behind today's CVEs.
›Prioritized by real exploit data
›Connect a repo in minutes
Run a free scan →
Live NVD · EPSS · CISA KEV