Backdoored AI Coding Extension: A Team Security Guide
A tampered Claude Code extension reached Cursor users on October 29, 2026. This guide explains what happened and how teams can audit extensions, respond to exposure, and govern AI coding tools.
The Short Answer
On October 29, 2026, a backdoored AI coding extension reached developers through a channel most teams treat as safe by default: the official extension marketplace inside their code editor. For about 11 hours, modified versions of the Claude Code extension were served through the Cursor VS Code Extension Marketplace, and Cursor later said at most…
On October 29, 2026, a backdoored AI coding extension reached developers through a channel most teams treat as safe by default: the official extension marketplace inside their code editor. For about 11 hours, modified versions of the Claude Code extension were served through the Cursor VS Code Extension Marketplace, and Cursor later said at most 5,000 users were affected. The modified builds were not developed by Anthropic, and Anthropic has said its own systems were not compromised. This guide explains what happened and what a responsible team should do next.
The uncomfortable lesson is simple. An editor extension is not an app you install and forget. It is code that runs where your source code, credentials, cloud sessions, and deployment keys live. When that code is tampered with, the attacker does not need to break into your servers first. They are already beside your codebase, with the same access your editor has. Treat every extension as unreviewed code running inside your codebase, and your approval process will look very different.
What happened on October 29, 2026
Cursor reported that malicious versions of the Anthropic.claude-code extension, versions 2.1.312 and 2.1.313, were available in its VS Code Extension Marketplace from about 2:27 AM to 1:35 PM Pacific Time on October 29, 2026. Distributed under the legitimate extension identity for Windows and macOS, they arrived looking like a routine update. Cursor removed the packages, temporarily shut down its marketplace, audited access, rotated keys, and rebuilt the marketplace before restoring service.
What remains unexplained is how the attacker gained the ability to publish a tampered build under a trusted publisher identity. Microsoft has confirmed that the malicious packages were never published through the official Visual Studio Marketplace or through npm. The same version numbers from Anthropic's own channels were not tampered with. For every team checking exposure, where the package came from matters as much as what it was called.
Cursor later updated its notice to add version 2.1.312 and correct its indicators of compromise. Teams that used the original notice should pull the indicators again and re-check their machines. One firm investigating linked incidents reported activity on a customer machine about eight hours after the exposure window closed: a short distribution window does not mean a short risk window.
What this backdoored AI coding extension actually did
According to analyses by Microsoft, Phorion, and OpenSourceMalware, the tampered package kept the legitimate code intact and added a malicious loader at the start of its main JavaScript file. The loader ran when the extension module loaded, before normal activation, with no AI conversation and no extra click required. The victim did not have to paste a malicious prompt or visit a suspicious site. Opening the editor was enough.
The loader collected host details including the hostname, username, operating system, architecture, home directory path, working directories, configured Git email address, and public IP address. It also gathered the names, not the values, of environment variables that looked like keys, tokens, or cloud credentials. It then tried to append an attacker controlled SSH public key, labelled svc-telemetry, to the user's authorized keys file, enable the local SSH service, and open a reverse SSH tunnel, retrying about every 60 seconds if the connection dropped.
Persistence was built for more than one platform. On macOS, the payload created a launch agent designed to look like an Apple component. On Windows, it created a scheduled task, with a fallback script in the Startup folder reported in some analyses. The package also contained code targeting Linux through a cron entry that runs at reboot. Cursor has noted that enabling local SSH services requires administrator privileges the extension lacked on most macOS and Windows hosts, so that step probably failed on many machines. Even so, persistence scripts outside the extension directory could survive removal of the extension, which is why uninstalling alone may not be sufficient.
Why extension marketplaces are a trust risk
Most developers install extensions the way consumers install phone apps: they search, recognise a trusted name, click install, and allow automatic updates. That habit made sense when extensions were small helpers for syntax highlighting. Modern AI coding extensions read files, run commands, call external services, and interact with repositories and cloud tools. A compromised extension therefore has a direct path to a team's most sensitive material.
The Cursor incident also breaks a comforting assumption about provenance. Teams often check the publisher name and version number and stop there. Here, the publisher identity and extension ID were legitimate, and the tampered build still reached users through an official marketplace. A name check cannot catch a publishing pipeline that has itself been abused. Trust must be verified at the level of the build, the distribution channel, and the machine, not only at the brand on the listing page.
Developer machines are unusually valuable targets because they bridge many systems at once. A single workstation may hold repository access, cloud credentials, staging and production keys, signing certificates, and chat and email sessions. A backdoor aimed at developers is therefore a potential route into code, infrastructure, and customer data. That is why cybersecurity for development teams must cover the editor and build environment, not only production servers.
How to audit for a backdoored AI coding extension in your team
Start with an inventory. Ask every developer and contractor to list the editor, marketplace source, and version of every AI coding extension installed on October 29, 2026, plus automatic updates received that day. Cross check that list against the exposure window: Cursor marketplace builds of anthropic.claude-code versions 2.1.312 and 2.1.313 served between about 2:27 AM and 1:35 PM Pacific Time. Record the source marketplace alongside the version number, since the same numbers from other channels were not affected.
Next, check the machine itself, not just the extension list. Look for an unexpected SSH key labelled svc-telemetry in the authorized keys file, review scheduled tasks, launch agents, and cron entries added around October 29, and review SSH logs and outbound connections. Treat any machine that ran the tampered build during the window as suspect until checked, even if the extension has since been updated or removed.
Use a structured checklist so nothing is skipped under pressure:
- Confirm the editor, marketplace source, extension name, and version for every developer machine, including laptops used at home.
- Check authorized keys files for the svc-telemetry entry and any other unfamiliar key.
- Inspect persistence locations: Windows scheduled tasks and Startup folder entries, macOS launch agents, and Linux cron entries.
- Review SSH and editor logs for activity around October 29 and in the days that followed.
- List every credential, token, and cloud session reachable from each suspect machine, and prepare to rotate them.
- Record findings per machine, with the person, date checked, and action taken.
Do not rely on a single tool to declare a machine clean. Endpoint protection, log review, and manual checks of persistence locations complement each other. In incidents investigated by one security firm, affected machines ran well known endpoint tools, which shows detection depends on looking in the right places.
Incident response if your team installed it
If a machine is confirmed or strongly suspected to have run the tampered build, treat it as compromised. First, preserve evidence: note the extension version, marketplace source, install or update time, and log entries before cleaning. Then remove persistence mechanisms separately from the extension itself, because uninstalling may leave the backdoor behind. Remove any unauthorized SSH key, malicious launch agent, scheduled task, or cron entry, and any fallback scripts.
Second, rotate credentials. Change every password, token, API key, and cloud credential accessible from the affected machine, and revoke sessions it could reach. Prioritise repository hosting accounts, cloud consoles, deployment keys, signing certificates, and service accounts the developer used. If a shared credential was reachable from that machine, assume it is no longer private.
Third, look outward. Review recent commits, pull requests, deployments, and infrastructure changes made from the affected machine. Check for new repository collaborators, new SSH keys on hosting accounts, new cloud users or keys, and unexpected outbound connections. If customer data or production systems could have been reached, escalate through your normal incident process immediately, and document every step.
Finally, reinstall only from a verified official channel. Anthropic published version 2.1.314 on Open VSX and the Microsoft Visual Studio Marketplace after the incident, so confirm the source channel rather than accepting an update prompt at face value.
Lock versions and govern AI coding assistants
The most practical defence against the next tampered update is deliberately boring: control what can be installed, pin what is approved, and review changes before they roll out to everyone. An allowlist for editor extensions, maintained by the team lead or security owner, turns installation from an individual habit into a team decision. Every listed extension should have a named owner, a business reason, a review date, and an approved marketplace source.
Version pinning deserves the same attention. Automatic updates are convenient, but they also mean a tampered release can reach every developer at once, as this incident showed. Where the editor supports it, pin approved extensions to reviewed versions and stage updates: one or two machines first, a short observation period, then a wider rollout. For AI coding assistants, define which tools are approved, which repositories and data they may access, and whether they may run commands or only suggest code.
Governance also means separating concerns. Developers should not run their daily editor with standing access to production secrets, and AI assistants should not be given credentials they do not need. Use short lived tokens and scoped repository access. When AI generated code enters a project, it still needs human review and testing before it ships, as described in this QA playbook for testing AI generated code, and a second safeguard is to quiz code before it merges, as described in this guide to a Git hook that quizzes developers on AI generated code.
Training completes the policy. Developers should know how to spot a suspicious extension update, how to report it without fear of blame, and what to do in the first hour after a suspected compromise. Attackers often reuse the confusion around supply chain incidents for phishing, so refresh awareness of deceptive messages with this AI powered phishing protection guide for 2026.
A practical extension governance policy for small teams
A one page policy, reviewed every quarter, is enough for most small and mid sized teams:
| Control | What to do |
| Allowlist | Approve extensions by name, publisher, version, and marketplace source. Block everything else by default. |
| Staged updates | Test extension updates on a small group first, then roll out to the wider team after a short observation period. |
| Least privilege | Give editors and AI assistants only the access they need for the current project. |
| Credential hygiene | Use short lived, scoped tokens. Never store production secrets in an editor profile. |
| Incident playbook | Write down who to call, what to check, and what to rotate if an extension is suspected to be compromised. |
| Regular review | Reconfirm the allowlist every quarter, remove unused extensions, and check for unreviewed installs. |
Review the policy after every incident, including incidents that did not affect you. The Cursor case is a useful tabletop exercise: walk through it with your team, ask who would have noticed a tampered update, and fix the gaps you find before a real incident tests them. The goal is not to ban AI coding tools. It is to use them with the same discipline you apply to production code, because that is exactly what they now touch.
Frequently Asked Questions
Was the Claude Code extension itself hacked?
No, according to Cursor and Anthropic. The malicious builds served through Cursor's marketplace were not developed by Anthropic, and Anthropic has said its systems were not compromised. The same version numbers from Anthropic's own channels were not affected.
Does uninstalling the extension remove the backdoor?
Not necessarily. Cursor has said uninstalling may not be sufficient, because persistence scripts and SSH keys written outside the extension directory can remain. Check authorized keys files and persistence locations separately, and remove any malicious entries before treating a machine as clean.
How do I know if my team was affected?
Check whether anyone installed or updated anthropic.claude-code from the Cursor marketplace between about 2:27 AM and 1:35 PM Pacific Time on October 29, 2026, specifically versions 2.1.312 and 2.1.313 for Windows and macOS. Then check each suspect machine for the svc-telemetry key and unfamiliar persistence entries.
Should teams stop using AI coding assistants after this incident?
Stopping outright is rarely realistic. The safer response is governance: an allowlist, pinned and staged updates, scoped access, credential hygiene, and human review of AI generated code.
What is the single most important change to make this week?
Build the extension inventory and audit machines that updated on October 29. Once you know what is installed, where it came from, and which version it is, every other control becomes straightforward to apply.
If your team needs help auditing developer machines or locking down AI coding tools, talk to our team through our contact page and we will help you review your setup and put the right controls in place.
Related Articles
Keep reading — more guides you may find useful.
Related Services
Services connected to this guide.
