Home>Spin.AI Blog>Browser Security>Claude Code Extension Hacked: Inside the Cursor Supply Chain Attack

Claude Code Extension Hacked: Inside the Cursor Supply Chain Attack

Nov 16, 2026 | Reading time 7 minutes
Author:
Melissa Garcia - Senior Marketing Manager

Senior Marketing Manager

On October 29, 2026, developers who installed or auto-updated Anthropic’s Claude Code extension in Cursor may have received a trojan instead.

For about 11 hours, Cursor’s extension marketplace served two tampered builds, versions 2.1.312 and 2.1.313, that tried to plant a persistent SSH backdoor on macOS and Windows machines. Up to 5,000 users may have pulled the malicious build.

Anthropic didn’t build these versions, and its own systems weren’t breached. The attacker never touched the real extension’s source. Instead, they got into the distribution channel that developers trusted to deliver it.

Anthropic told SecurityWeek, “A malicious actor used access to Cursor’s marketplace to publish a tampered copy of the Claude Code extension, which we didn’t build or publish.”

That makes this one of the clearest examples yet of an extension supply chain attack, and a warning for every organization that lets employees install browser extensions freely, whether in a code editor or a web browser.

What happened: the Claude Code extension hack at a glance

Cursor is a popular AI code editor built on VS Code, and it runs its own extension marketplace. According to Cursor’s security notice, tampered copies of the anthropic.claude-code extension were live there from about 2:27 a.m. to 1:35 p.m. PT on October 29. The extension kept working normally, so users had little reason to suspect anything.

When (2026, PT)What happened
Before Oct 29An attacker uses a Cursor access key leaked in the Fireworks AI breach, moves through Cursor’s SaaS tools, and steals a credential for the cache that serves the extension marketplace
Oct 29, ~2:27 a.m.Trojanized Claude Code versions 2.1.312 and 2.1.313 for macOS and Windows appear on the marketplace cache
Oct 29, 2:27 a.m. – 1:35 p.m.Up to 5,000 users who install or auto-update receive the malicious build. On load, it profiles the host, plants an SSH key, tries to open a reverse tunnel, and installs persistence scripts
Oct 29, ~1:35 p.m.Cursor pulls both versions and takes the entire marketplace offline
Oct 29–30Cursor audits access and finds no other tampered extensions. It rotates keys, rebuilds and restores the marketplace, and gets the attacker’s server taken offline
Oct 30, 11:30 a.m.Cursor updates its indicators of compromise, confirms the affected versions, and offers detection and cleanup scripts
Early NovAnthropic ships a clean version, 2.1.314, so users can easily tell it apart from the bad builds
Nov 1Last attacker activity Cursor detected
Mid-NovCursor names the stolen Fireworks AI key as the root cause and notifies affected users. It also moves its inference traffic off Fireworks

Microsoft confirmed the malicious package never reached the Visual Studio Marketplace, Open VSX, or npm.

How it happened: from a stolen API key to a trojanized extension

The attack didn’t start at Cursor. It started with a breach at Fireworks AI, an inference provider Cursor worked with. Cursor’s disclosure lays out the chain:

  • A third-party key leaks → An attacker obtained an access key Cursor had issued to Fireworks AI, exposed in Fireworks’ own breach.
  • Shared data is copied → The key unlocked a read-only storage bucket Cursor used to share data with Fireworks. It held coding agent traces, mostly from 2025, from sessions not in Privacy Mode. Some traces contained unredacted credentials, exposing API credentials for under 0.01% of customers.
  • The attacker moves laterally → Credentials found in that data gave access to select SaaS tools Cursor uses, and then to more internal data.
  • The marketplace is poisoned → Using another stolen credential, the attacker placed trojanized Claude Code builds on the cache that serves Cursor’s marketplace.

What the malicious extension did

Microsoft found an injected JavaScript loader that runs as soon as the extension loads, before any legitimate code. It used built-in Node.js APIs and system tools, so no second download was needed. It also hid its errors and output. The loader then:

  • Profiled the machine: hostname, username, OS, home and working directories, Git email, public IP, and the names of environment variables containing words like KEY, TOKEN, SECRET, or AWS.
  • Planted an SSH key: it added an attacker key, tagged svc-telemetry, to authorized_keys and tried to turn on the local SSH server.
  • Opened a reverse tunnel: an outbound SSH connection to 134.209.82.225 that would expose the workstation’s port 22 to the attacker, retried every 60 seconds.
  • Set up persistence: a fake com.apple.sysmond LaunchAgent on macOS, a SysMon scheduled task or Startup script on Windows, and a cron @reboot entry on Linux.

Cursor says enabling SSH needs admin rights the extension lacked on most hosts, so the tunnel usually failed. But the key, the persistence scripts, and the data theft could still succeed.

Cursor warned, “Uninstalling the extension may not be sufficient – the payload drops standalone persistence scripts that keep running.”

Why this incident matters

This attack was significant because it broke the trust users place in a brand they know.

Nobody was tricked into installing a lookalike. Developers installed the official Claude Code extension from the marketplace their editor points to by default, and auto-updates did the rest.

Three lessons stand out:

  1. Third-party risk compounds. One leaked vendor key led to a data copy, then SaaS access, then control of a software distribution channel. The same Fireworks AI breach also hit Uber and Notion, per The Hacker News.
  2. Developers are high-value targets. Their machines hold SSH keys, cloud credentials, and source code. The payload’s hunt for variable names like AWS, GCP, and TOKEN shows exactly what the attacker wanted next.
  3. Takedowns don’t clean endpoints. Cursor removed the builds in hours, but devices already infected kept their backdoor scripts. Security teams needed an inventory of who had installed what, and when, to respond at all.

The bigger risk: browser and IDE extensions

Code editor extensions and browser extensions share the same weak spot. Both run with broad access, both install with a click, and both update silently in the background.

A clean extension today can turn malicious tomorrow through one compromised update, a sold developer account, or a breached marketplace.

Browser extensions and malicious browser extensions are the larger exposure for most companies, because every employee has a browser. A risky extension can:

  • Read and change data on every site a user visits, including SaaS apps like Google Workspace, Microsoft 365, and Salesforce
  • Steal session cookies and tokens to take over accounts without a password
  • Copy sensitive data into unsanctioned GenAI tools
  • Inject phishing content into trusted pages

Most organizations can’t answer a basic question: which extensions are installed across our users, and which version? When an incident like this one breaks, that blind spot turns a few-hour exposure into a long investigation.

Lessons learned: how to reduce extension risk

The Claude Code extension hack points to a few practical steps any security team can take now:

  • Allow only approved extensions. Keep an allowlist of vetted extensions and publishers, and block everything else by default.
  • Don’t trust updates blindly. A safe extension can turn malicious in one release. Review each new version, or delay updates until a version has a track record.
  • Know what’s installed, everywhere. Keep a live inventory of every extension, version, user, and device so you can scope an incident in minutes.
  • Watch for risky permissions. Flag extensions that can read every site, access cookies, or send data to outside services.
  • Plan for cleanup beyond uninstall. Malicious extensions can leave persistence behind, so pair removal with endpoint checks and credential rotation.
  • Treat vendor access as attack surface. Limit and rotate the keys you share with third parties, and keep secrets out of logs and shared data.

How SpinCRX protects against extension supply chain attacks

These key lessons are critical; but impossible to do manually and at scale.

SpinCRX turns those lessons into automated policy for the browser, where most extension risk lives. It finds every extension in your environment, scores its risk, and blocks or removes dangerous ones before they can steal data.

What you needHow SpinCRX does it
Approved extensions onlyPolicy-based allowlists and blocklists, plus an approval workflow that shows risk details for every install request
Safe updatesEvery extension version is risk-assessed separately, so a risky new release is flagged even when earlier versions were safe
Full visibilityOne inventory of extensions across Chrome, Edge, Firefox, and Safari, on every profile and device
Risky permission detectionAI-driven scoring of permissions, publisher reputation, sandbox behavior, and whether an extension can send data to external LLMs
Fast responseAutomated policies block or revoke risky extensions in seconds, not days
No blind spotsThe endpoint agent covers all browser profiles on managed devices, including personal ones; an agentless option covers BYOD

SpinCRX’s risk engine has already assessed more than 1 million browser extensions and apps. Monitoring runs 24/7, and integrations with CrowdStrike, Splunk, and ServiceNow send findings into the tools your SOC already uses.

Take control of your extensions before the next attack

The Claude Code extension hack won’t be the last extension supply chain attack.

The next one may arrive through a Chrome extension your sales team installed two years ago, or an extension your intern adds to their workspace within Safari because they want to experiment with AI.

You can’t patch that kind of risk after the fact. You have to see it and govern it before the bad update lands.

SpinCRX gives you that control without forcing users to switch browsers. Deploy it with an endpoint agent, an agentless browser extension for BYOD, or both.

Sources

Was this helpful?
Melissa Garcia - Senior Marketing Manager

Written by

Senior Marketing Manager at Spin.AI

Melissa Garcia is the Senior Marketing Manager at Spin.AI, responsible for driving brand strategy, content marketing, product positioning, and integrated marketing initiatives across the company's cybersecurity portfolio.

With a background in B2B SaaS and cybersecurity marketing, Melissa specializes in building compelling brand narratives, translating complex technologies into engaging stories, and developing strategic content strategies that builds brands, fuels company growth, and connects products with their audiences.

Prior to joining Spin.AI, Melissa led marketing at DoControl, helping establish the company's market presence and positioning in the SaaS security space.

Her primary focus is brand positioning, product marketing, content strategy, thought leadership, and storytelling that helps organizations understand the value of emerging cybersecurity technologies.