Cursor investigation finds about 5,000 users may have installed the backdoored Claude Code extension
Cursor said on Nov 13 that malicious Claude Code extension builds were live about 11 hours on Oct 29, reaching up to 5,000 users. Privacy Mode users were safe.

Cursor published its investigation into the malicious Claude Code extension on November 13, 2026: tampered builds of Anthropic.claude-code 2.1.312 and 2.1.313 were live on its extension marketplace for roughly 11 hours on October 29 — from 2:27 a.m. to 1:35 p.m. Pacific time — and up to 5,000 users who requested those versions in that window may have installed the malicious build. The headline numbers: Privacy Mode users’ prompts and code were not affected; among users without Privacy Mode, less than 0.05% had coding-agent traces exposed and less than 0.01% of customers had un-redacted API credentials leak. What makes the disclosure worth reading is the attack chain’s starting point — not Cursor’s own infrastructure, but an access key Cursor had issued to its inference vendor, Fireworks.
Key points
- The window: malicious versions 2.1.312 and 2.1.313 sat on the cache serving Cursor’s extension marketplace from 2:27 a.m. to 1:35 p.m. PT on October 29, about 11 hours, reaching at most 5,000 users.
- The exposure: under 0.05% of non-Privacy-Mode users had agent traces copied, under 0.01% of customers had un-redacted credentials leak; Cursor last observed attacker activity on November 1 and considers the incident contained.
- The root cause: an attacker used a Fireworks-leaked access key Cursor had issued to reach a shared read-only storage bucket, pivoted through un-redacted credentials found in traces into internal SaaS tools, then planted the backdoored extension on the marketplace cache with a stolen credential.
Background
The incident went public on October 29, when Cursor found the tampered extension on its marketplace cache and published its first security notice the same day; a follow-up on October 30 added indicators of compromise, detection scripts and remediation steps, and affected users and team admins were notified through in-app notices and email. The impersonated party was Anthropic’s Claude Code — Anthropic never published those two builds, the official distribution channels and the Open VSX upstream marketplace never carried the malware, and the weakness sat in the cache layer of Cursor’s own marketplace.
The first link of the chain was outside Cursor. The investigation confirms the attacker obtained an access key Cursor had issued to Fireworks, and, according to Cursor, Fireworks’ own security bulletin acknowledges the key was exposed on its side. Fireworks has had a rough stretch: October’s fallout included Notion AI transcripts routed through it, and on November 9 it issued another bulletin covering billing data and incomplete model weights. One leaked vendor key ended up reaching an AI coding tool’s extension distribution pipeline — that chain of transmission is the part of this story worth remembering. We covered a different leak in the same toolchain in Claude Code caught sending account emails in User-Agent headers: data that should not leave keeps finding a way out.
The facts
Cursor’s November 13 update puts hard numbers on every layer of the incident, consistent across the official post and secondary reporting from IT之家 and others.
| Item | Official figure |
|---|---|
| Malicious versions | Anthropic.claude-code 2.1.312, 2.1.313 |
| Live window | Oct 29, 2026, 02:27–13:35 PT, about 11 hours |
| Users possibly affected | Up to about 5,000 (requested those versions on Oct 29) |
| Agent trace data | Under 0.05% of non-Privacy-Mode users exposed |
| Un-redacted credentials | Under 0.01% of customers affected |
| Last attacker activity | Nov 1, 2026, blocked |
The inventory of what leaked is equally specific: the attacker copied coding-agent traces from the shared read-only bucket — mostly 2025 sessions from non-Privacy-Mode users whose data was permitted for model training — some of which contained un-redacted credentials. Using those, the attacker reached a set of SaaS tools and obtained archived non-production source repositories (with business contact information), limited customer support data and emails, product usage metrics, summaries of some coding-agent conversations, and credentials users had submitted to support. Cursor’s response moved all inference traffic off Fireworks and deactivated its key, rotated every exposed credential, rebuilt the marketplace, tightened network access restrictions, added secret scanning across internal data sources, redeployed production infrastructure as a precaution, and engaged external forensics. Officially unaffected: live traffic, production systems, the Cursor app and its shipped code, other marketplace extensions, and customer code repositories.
What others say
IT之家’s November 14 report relays the official numbers accurately and stresses the notice’s warning that uninstalling the extension alone does not remove the risk. On the analysis side, ExplainX frames the incident as “breach chaining through AI suppliers”: one vendor’s leaked key connected Notion AI transcripts, October’s wider fallout and Cursor’s marketplace cache into a single story — any link on the subprocessor list failing widens every customer’s exposure. That analysis also cites third-party teardown work on 2.1.313 showing the payload drops standalone persistence on the host, which is exactly what Cursor’s notice warned about. One detail worth flagging: the advice to grep ~/.ssh/authorized_keys for an SSH key labeled svc-telemetry, along with the official detection and remediation scripts, comes from Cursor’s original October 29 notice — the IT之家 report does not reproduce it, so you need the primary post to get those steps.
Checks and cleanup
Cursor’s guidance for affected users is blunt: if you installed or updated to 2.1.312 or 2.1.313 during the window, follow the October 30 notice end to end — the post states plainly that “uninstalling the extension may not be sufficient,” because the payload drops standalone persistence scripts. For everyone else, three checks cover the essentials: look at installed extension versions against the malicious range, confirm where your extensions come from rather than trusting a marketplace cache’s history, and search ~/.ssh/authorized_keys for the label svc-telemetry — finding it means the machine is impacted. Cursor emailed every user whose data was directly exposed, and team admins of affected customers received a specific list of keys to rotate; if your team ever pasted API credentials into a support ticket, rotating them now is the right call regardless.
Our take
What this incident changes is not one tool’s security rating but the assumption that a marketplace equals its upstream. Developers assume a marketplace listing mirrors the official package; Cursor’s breach shows that when the marketplace is a self-run cache, that equality needs checking — glancing at an extension’s source and version before installing should not be a publisher-only habit. Vendor risk got repriced too: a Fireworks key leak cascaded into a customer’s extension marketplace, so weighing key-management practices when choosing an inference provider is now due diligence, not paranoia. The honest remaining question: users without Privacy Mode who updated in those 11 hours are only now learning whether they were part of that 0.05%. For most people it was a near miss, but for the ecosystem it is one more piece of hard evidence — after PixelLeak — for the question of where agent data actually goes.