Update, 31 October 2026: Microsoft Threat Intelligence published its analysis of the loader on 30 October. Cursor and Microsoft now name two tampered versions, 2.1.312 and 2.1.313, not 2.1.313 alone.
If you installed or updated the Claude Code extension in Cursor on 29 October 2026, check your machine today. Cursor says modified builds, Anthropic.claude-code versions 2.1.312 and 2.1.313, sat in its extension marketplace from about 2:27am to 1:35pm Pacific time. Cursor says the builds were not developed by Anthropic. It says at most 5,000 users who requested specific cached versions potentially received a malicious build; this is not a confirmed infection count.
The builds try to plant a persistent SSH backdoor, and Cursor warns that uninstalling the extension may not remove it. This playbook gives the checks and the cleanup in the order to run them, then the controls that stop the next malicious extension.
Malicious Extensions: How Cursor Linked the Incident Chain
Investigation update, 13 November 2026: Cursor says an attacker used a key it had issued to Fireworks after Fireworks exposed it. The key allowed a copy of a read-only dataset of historical coding agent traces, mostly from 2025 sessions outside Privacy Mode. Some traces held unredacted credentials; Cursor says fewer than 0.01% of customers had API credentials exposed.
Cursor says those credentials led to select SaaS tools it used, then to limited support and business data and archived non-production source code. A credential taken from an internal system let the attacker place unauthorized Claude Code extension builds in Cursor’s marketplace cache on 29 October. Cursor says the source marketplace, Open VSX, did not hold the malicious builds. This reported chain can inform cyber threat intelligence triage, but it remains Cursor’s account rather than independently established attribution.
Scope of Cursor’s estimate: Cursor’s upper bound of 5,000 means users who requested specific cached versions and potentially received a malicious build. It is not a count of confirmed installations or infections. Cursor says it blocked the last attacker activity it observed on 1 November and believes the incident is contained.
What the New Findings Change for Response
Cursor says it directly contacted affected customers. Team admins whose API credentials were exposed received a list of keys to rotate; Cursor also asked the issuing services to revoke those credentials. Customers whose support or business data may have been involved received separate notices. A notice about a credential or support record calls for that tailored response, even if no tampered extension was installed.
The 29 October exposure window applies to Mac and Windows hosts that installed or updated Claude Code 2.1.312 or 2.1.313 from Cursor’s marketplace. For those hosts, follow Cursor’s host remediation notice and the playbook below. Check for execution and persistence, isolate an affected host, then clean or rebuild it and rotate reachable secrets from a trusted machine. Cursor warns that uninstalling alone may leave persistence behind.
Brandefense Blackhole counted 13,507 dark-chat posts matching the broad term “Cursor” from 16 August to 14 November 2026. The measured unit is posts mentioning that term, not incident posts or compromised users. This volume cannot establish how many people received the malicious build.
What Did the Tampered Builds Do?
Cursor’s security notice says the modified extension ran code when loaded and attempted to install a persistent SSH backdoor. It collected host and user context, attempted to enable remote access, and tried to keep a tunnel available after the editor session ended. These are reported behaviors of the malicious build, not proof that every step succeeded on every machine.
Cursor says enabling the local SSH service usually requires administrator privileges that the extension did not have on macOS and Windows. It also worked with the hosting provider to take the attacker’s server offline. Those limits reduce one route to access, but Cursor still recommends remediation because the malicious build could leave changes outside the extension.
Which Marketplace Served the Bad Build?
Early reports confused the distribution point. Cursor says the unauthorized copies were placed directly in its marketplace cache; the Open VSX source marketplace did not hold the malicious builds.
Our own registry check on 30 October at 09:03 UTC matched that. Open VSX and the Visual Studio Marketplace list versions 2.1.312 and 2.1.314 of anthropic.claude-code, and neither lists 2.1.313. The 2.1.312 there is the clean release; Cursor’s notice names a tampered 2.1.312 in its own marketplace.
StepSecurity first wrote that 2.1.313 came from Open VSX. On 30 October it corrected its post, citing Anthropic that the build was never published to Open VSX or the Microsoft marketplace. Microsoft says the packages were never distributed from the Visual Studio Marketplace or npm, and all evidence points to Cursor’s marketplace alone. Anthropic confirmed to Microsoft that its systems were not compromised.
The lesson is that one extension ID can live in several registries. A trusted publisher name in one store says nothing about the copy another store serves. Model hubs have the same gap: in the Hugging Face malware campaign, a familiar Qwen2 name hid Python that ran on load.
The Playbook: Find, Contain, Clean, Rotate
Use Cursor’s 29 October security notice for hosts that installed or updated Anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace during its stated exposure window. Review the extension source and version, then correlate installation, endpoint and network telemetry. Cursor provides its own validation and remediation resources in that notice.
- Identify potentially exposed hostsReview extension-installation records for the affected versions and the Cursor marketplace exposure window.
- Investigate executionCorrelate endpoint and network telemetry with Cursor’s notice; treat a match as an investigative lead.
- Isolate affected hostsPreserve evidence and disconnect hosts showing execution or backdoor activity before cleanup.
- Clean or rebuildRemove the unauthorized access and persistence mechanisms, then assess whether rebuilding the host is safer. Uninstalling the extension alone may be insufficient.
- Rotate from a trusted systemRotate credentials and sessions accessible from the affected host after containment and cleanup.
The order follows Cursor’s “What you should do” guidance. It separates potential installation from confirmed host impact and keeps credential rotation away from a host that may still be compromised.
How Does Microsoft Defender Detect It?
Microsoft Threat Intelligence describes Defender alerts and hunting guidance for suspicious activity associated with a code editor, unauthorized remote access and persistence. Use those signals alongside extension-installation and network telemetry to scope an investigation.
Microsoft cautions that a hunting result is an investigative lead rather than proof of compromise. Confirm a finding against the host’s installation history and behavior, then follow Cursor’s isolation and remediation guidance for affected machines.
Why Run the Steps in This Order?
Cursor says uninstalling the extension may leave standalone persistence behind. Start with extension and host evidence, then isolate a host showing malicious execution before removing unauthorized access and persistence. Assess whether a rebuild is needed rather than assuming an uninstall restored trust.
Rotate reachable credentials and sessions from a trusted system after containment. This also covers credentials that may have been accessible later through the backdoor, beyond the limited environment information the build initially collected. Our secrets rotation checklist from the Railway breach gives a broader order for cloud and CI keys.
How Do You Stop the Next Malicious Extension?
This is not the first time an editor extension became the entry point. The GlassWorm malware spread through VS Code and Open VSX extensions. The Nx Console extension breach ended with GitHub-internal repositories stolen, and our questions on dev tools as an entry path still apply. Both fit the MITRE technique for compromising a software supply chain.
A different misuse of the same agentic tools runs the intrusion itself. Our defender playbook for Aur0ra’s AI agent intrusions covers a ransomware group that used an AI coding agent for reconnaissance, credential attacks and lateral movement.
Four controls reduce the risk without banning extensions. The same ideas apply to browser extensions, which one vendor report says a campaign it calls ShadyPanda abused. A researcher also describes extensions steering Claude Desktop; see the Claude Dispatch flaw explained.
- Allow-list by publisher, ID and version. VS Code 1.96 and later has an extensions.allowed setting, and admins can enforce it as the AllowedExtensions policy. It can pin an extension to listed versions and platforms.
- Know which registry each editor uses. Cursor served these unauthorized builds from its own marketplace cache. Check each editor’s registry and policy controls; VS Code settings may not apply to a fork.
- Delay brand-new versions. Microsoft points Cursor users to Cursor’s hardening guidance: allow only approved publishers and set an installation cooldown. A cooldown holds back versions published only hours earlier.
- Treat developer laptops as privileged endpoints. Keep long-lived cloud keys out of shell profiles, and alert on unusual editor activity, remote access and persistence changes.
Engineer laptops with production access were the weak point in our DevOps security lessons from October 2026 too. An extension that runs on load is one more automated system with standing access to those machines. An editor plugin can also widen an existing flaw’s reach: see how the optional NppExec plugin figures in Notepad++ 8.9.9’s plugin command execution fix. The scheduled task trick here is scheduled task persistence, so existing detections for that technique apply.
A useful first move this week: export the extension list from every developer machine, with versions and the registry each came from. Anything you cannot explain goes on the allow-list review. A separate Claude Code issue lets an approved permission file leak secrets; see how the Don’t Ask Again setting exposes them.
For another software distribution compromise, the WPPConnect malicious npm and Docker releases show why defenders should verify deployed package versions and container digests.
Frequently Asked Questions
Was the Claude Code extension on the Microsoft Marketplace affected?
No. Microsoft says the malicious packages were never published to or distributed from the Visual Studio Marketplace or npm. OpenSourceMalware also lists the Microsoft Marketplace extension as not malicious. On 30 October, that marketplace listed no version 2.1.313.
Does uninstalling the extension remove the backdoor?
Not on its own. Cursor says the payload may leave standalone persistence after the extension is uninstalled. Investigate and isolate an affected host, remove unauthorized access and persistence or assess a rebuild, then rotate reachable credentials from a trusted system.
Is the Claude Code npm package affected?
No. OpenSourceMalware says the @anthropic-ai/claude-code npm package is not malicious, and Microsoft says npm was not affected. The modified code was in editor extension builds 2.1.312 and 2.1.313 served by Cursor's marketplace.
Does Microsoft Defender detect the malicious extension?
Microsoft describes Defender alerts and hunting guidance for suspicious editor activity, remote access and persistence. A match is an investigative lead, not proof of compromise. Correlate it with extension-installation and host telemetry before following Cursor's remediation guidance.
Sources
- OpenSourceMalware: Trojanized Claude Code Extension Plants an SSH Backdoor
- Cursor: Security notice on Anthropic.claude-code 2.1.312 and 2.1.313 in the Cursor Extension Marketplace
- Visual Studio Code: Manage extensions in enterprise environments
- Visual Studio Code: Command line interface
- Cursor: An update on our recent security incident (13 November 2026)
- StepSecurity: Trojanized Claude Code Extension Delivered Persistent SSH Backdoor to Developer Machines
- Microsoft Security Blog: Inside a malicious Claude Code-branded extension and its persistent SSH backdoor
