Threat level Guarded
Submit a tip

Major threats, straight to your inbox

Only the threats that matter: actively exploited flaws, big breaches and attacks that need action. No spam, unsubscribe any time.

Cursor Claude Code Extension Backdoor Traced to Stolen Fireworks Key

Cursor says an access key it issued to Fireworks AI, exposed in the Fireworks breach, let an attacker harvest credentials from agent traces and plant trojanized Anthropic.claude-code 2.1.312 and 2.1.313 builds that add an SSH backdoor. Up to 5,000 users may have received them.

Cursor Claude Code Extension Backdoor Traced to Stolen Fireworks Key

Threat card

Threat type
Supply Chain
Disclosed
Oct 29, 2026
Affected
Anthropic.claude-code 2.1.312 and 2.1.313 (Windows, macOS) served from the Cursor marketplace cache on 29 Oct 2026; Cursor coding agent traces and some customer API credentials

The short version

  • An access key Cursor issued to Fireworks AI, exposed in the Fireworks breach, was the entry point.
  • The attacker harvested unredacted credentials from non-Privacy Mode coding agent traces and reached Cursor SaaS tools.
  • Trojanized Anthropic.claude-code 2.1.312 and 2.1.313 builds sat in the Cursor marketplace cache for about 11 hours, reaching at most 5,000 users.
  • The payload adds an svc-telemetry SSH key and a persistent reverse SSH tunnel, so uninstalling the extension alone may not be enough.

Cursor has traced the Cursor Claude Code extension backdoor to an access key it had issued to inference provider Fireworks AI, which Fireworks exposed in its own breach. Using that key and credentials lifted from coding agent traces, the attacker planted trojanized builds of the Anthropic.claude-code extension, versions 2.1.312 and 2.1.313, in the Cursor marketplace cache on 29 October 2026. Up to 5,000 users may have received them during an 11 hour window. Anyone who installed or updated the extension in that window should check for an attacker SSH key now, because uninstalling alone may not remove the backdoor.

What happened

An attacker used a Cursor access key exposed in the Fireworks AI breach to copy coding agent traces, harvested credentials from them, reached Cursor’s SaaS tooling and pushed malicious Claude Code extension builds into the cache behind Cursor’s VS Code extension marketplace. The builds added an SSH key and a persistent reverse SSH tunnel on Windows and macOS hosts.

Cursor first described the malicious builds in its security notice on the Anthropic.claude-code versions, and on 13 November 2026 published an update tracing the incident to the Fireworks key. The key opened a read-only storage bucket Cursor used to share data with Fireworks. That bucket held coding agent traces, mostly from 2025, from sessions that were not in Privacy Mode. Some traces contained unredacted credentials, and fewer than 0.01% of customers had API credentials exposed.

This is a material development on two stories we have already covered: the Fireworks AI breach that exposed Notion AI chats and the Uber driver data exposure tied to the same Fireworks intrusion. Cursor is the third downstream victim we have tracked, and the first where the attacker turned stolen access into a software supply chain attack.

Timeline

Date Event
2025 (mostly) Coding agent traces from non-Privacy Mode sessions accumulate in the bucket Cursor shares with Fireworks AI.
Before 29 Oct 2026 The attacker uses the exposed Fireworks key to copy trace data, extracts credentials and reaches Cursor SaaS tools.
29 Oct 2026, about 02:27 PT Malicious Anthropic.claude-code 2.1.312 and 2.1.313 builds for Windows and macOS become available in the Cursor marketplace cache.
29 Oct 2026, about 13:35 PT Cursor removes the builds and publishes its security notice; it later reports the attacker server offline.
30 Oct 2026 Notice updated with indicators of compromise, check commands and remediation scripts; affected users and admins notified.
1 Nov 2026 Last attacker activity Cursor observed, which was blocked.
13 Nov 2026 Cursor publishes its incident update linking the entry point to the key it issued to Fireworks AI.

How the attack works

Initial access: a vendor key outside the vendor’s control

The entry point was not a Cursor vulnerability. It was a legitimate access key that Cursor had issued to Fireworks AI, which Fireworks exposed. Cursor’s incident update says the key gave read access to a storage bucket used for data sharing with Fireworks. Cursor has since deactivated the key, moved all inference off Fireworks and cut off its data access.

Credential access: secrets inside agent traces

The copied data was coding agent traces from sessions outside Privacy Mode, and some of those traces held unredacted credentials. Cursor says trace exposure affected fewer than 0.05% of non-Privacy Mode users, who were notified directly.

Lateral movement into SaaS tooling

With the harvested credentials the attacker reached Cursor SaaS tools. Cursor lists the data touched: archived non-production source repositories, limited customer support data and emails, product usage metrics, summaries of some agent conversations and credentials users had sent to support. Cursor says production systems, the Cursor app, official updates, release signing keys, customer code repositories and Privacy Mode prompts and code were not affected.

Supply chain impact: poisoning the extension cache

On 29 October the attacker used a stolen internal credential to place tampered copies of Anthropic.claude-code 2.1.312 and 2.1.313 into the cache that serves Cursor’s VS Code extension marketplace. Anthropic did not publish these builds, and Cursor says Open VSX never carried the malware. Cursor puts the maximum exposure at 5,000 users who requested the cached versions.

Execution and persistence on developer machines

According to the Cursor security notice, the payload in extension.js collects host and user details and sends them to the attacker, adds an SSH key labelled svc-telemetry to authorized_keys, enables the SSH service and opens a reverse SSH tunnel. It persists through a LaunchAgent on macOS and a scheduled task with a Startup folder fallback on Windows, retrying about every 60 seconds. Cursor notes the tunnel usually fails without administrator rights, but the key and the persistence scripts survive an extension uninstall.

MITRE ATT&CK mapping

Tactic Technique (ID) How it was used here
Initial Access Valid Accounts (T1078) Access key Cursor issued to Fireworks AI, exposed in the Fireworks breach.
Collection Data from Cloud Storage (T1530) Copied coding agent traces from the shared read-only bucket.
Credential Access Unsecured Credentials (T1552) Harvested unredacted credentials embedded in agent traces.
Initial Access Compromise Software Supply Chain (T1195.002) Trojanized Anthropic.claude-code builds placed in the marketplace cache.
Discovery System Information Discovery (T1082) Payload collects host and user details and sends them to the attacker.
Persistence Account Manipulation: SSH Authorized Keys (T1098.004) Adds the svc-telemetry key to authorized_keys.
Persistence Create or Modify System Process: Launch Agent (T1543.001) ~/Library/LaunchAgents/com.apple.sysmond.plist on macOS.
Persistence Scheduled Task/Job (T1053.005, T1053.003) Windows task SysMon; Linux @reboot cron entry.
Persistence Boot or Logon Autostart: Startup Folder (T1547.001) Windows fallback sysmon.cmd in the Startup folder.
Defense Evasion Masquerading: Match Legitimate Name (T1036.005) Persistence named com.apple.sysmond and SysMon.
Command and Control Protocol Tunneling (T1572) Reverse SSH tunnel to the attacker host on TCP 22.

Who is behind it

Cursor has not attributed the attack to a named actor. The only published infrastructure is a single IP address used for the beacon and the reverse tunnel, which Cursor says has been taken offline. The access path runs through the Fireworks AI intrusion, so defenders should read this alongside Fireworks’ own security bulletins rather than treat it as a separate campaign.

Affected products and versions

Product Affected Fixed CVE CVSS
Anthropic.claude-code via Cursor marketplace cache (darwin-arm64, win32-x64) 2.1.312, 2.1.313 served 29 Oct 2026, about 02:27 to 13:35 PT Builds removed; marketplace validated and rebuilt None assigned Not applicable
Cursor customer data (non-Privacy Mode) Coding agent traces, mostly 2025; API credentials for under 0.01% of customers Credentials rotated by Cursor None Not applicable

Not affected, per Cursor: the official Anthropic releases, the VS Code Marketplace, Open VSX, other marketplace extensions, the Cursor app and its signing keys, and Privacy Mode data.

Indicators of compromise

Type Indicator Context
SHA-256 fcbb327e38e61c4fec795e0d0591f52cd1014b813d20a5912104953c3a17c364 VSIX 2.1.312 darwin-arm64
SHA-256 3de1722901165769893bef57c72594e9a54459faa23f3091b136b79966edd081 VSIX 2.1.312 win32-x64
SHA-256 80402fdcad8799285fd93044976973ca000da7995fb93363f819bb8e4f84ee6c VSIX 2.1.313 darwin-arm64
SHA-256 6786ec90cbd45949695bd7805c9fbb26cfc91bc5fc897de0183d9bcb9cb21731 VSIX 2.1.313 win32-x64
SHA-256 ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15 Payload extension.js
IP 134[.]209[.]82[.]225 Reverse SSH (tunnel@, TCP 22) and beacon
URL hxxp://134[.]209[.]82[.]225:8099/beacon HTTP POST beacon
SSH key SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk Unauthorized key, comment svc-telemetry
SSH key SHA256:KOAU57+ZG6u3MG4WwzPzPX2O5YEe9qdOC1Bl+6MVyDQ Reverse tunnel key
File ~/.config/.smt, ~/.config/.smtk, $TMPDIR/.stk_<10 hex> macOS and Linux persistence
File %USERPROFILE%\.config\smt.cmd, %USERPROFILE%\.config\smtk Windows persistence
Task / plist SysMon, com.apple.sysmond Windows scheduled task; macOS LaunchAgent label

Cursor stresses that traffic to ifconfig[.]me or api[.]ipify[.]org on its own is not an indicator, since both are legitimate services the payload also queries.

Detection and hunting

Start with the cheapest check: search every developer endpoint’s authorized_keys files for the string svc-telemetry. On Windows that includes C:\ProgramData\ssh\administrators_authorized_keys. Cursor’s notice publishes one-line macOS and PowerShell checks that do exactly this.

Then hunt in process telemetry for ssh -N -R '*:0:localhost:22' or autossh -M 0 with StrictHostKeyChecking=no and ExitOnForwardFailure=yes. On Windows, look for schtasks /create /tn SysMon (Security event 4698 records task creation where auditing is enabled) and Add-WindowsCapability -Online -Name OpenSSH.Server~~~~0.0.1.0. On macOS, look for systemsetup -setremotelogin on and a new LaunchAgent named com.apple.sysmond.

In network logs, flag any outbound TCP 22 or TCP 8099 connection to the attacker IP, and loopback POSTs to 127[.]0[.]0[.]1:8099/beacon. In extension inventory, query for Anthropic.claude-code at 2.1.312 or 2.1.313 sourced from the Cursor marketplace, and match VSIX hashes from the table above.

Mitigation and response

  1. Identify every host that installed or updated Anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace on 29 October.
  2. Run Cursor’s check commands for the svc-telemetry key; isolate any host where it is present or where the payload ran.
  3. Remove the key, the LaunchAgent, scheduled task, Startup entry, cron entry and tunnel processes using Cursor’s remediation scripts; consider rebuilding the host.
  4. Rotate every credential and session reachable from an affected host, from a separate trusted system.
  5. Review which secrets your developers have pasted into AI coding agents, and turn on Privacy Mode or equivalent data controls where policy allows.
  6. Audit access keys you have issued to AI and inference vendors, scope them to the minimum data and restrict them by network where possible.

Analyst take

The Fireworks AI incident now reads as a hub breach. One inference provider’s exposure has produced confirmed downstream impact at Notion, Uber and Cursor, and at Cursor it escalated from data access to code execution on developer laptops. The weak point was not the model or the editor. It was a long-lived vendor key with read access to agent traces full of secrets.

Security teams should treat AI agent transcripts as credential stores, because in practice they are. Any organisation that shares traces, logs or prompts with a model provider should know exactly which keys grant that access, what those traces contain and how fast they can be revoked. Developer workstations with extension auto-update deserve the same scrutiny as build servers.

What to do now

0 of 5 done

Questions readers ask

Was the official Claude Code extension compromised?

No. Cursor says Anthropic did not publish the malicious builds, and the VS Code Marketplace, Open VSX and npm were not affected; only builds served from the Cursor marketplace cache were tampered with.

How do I know if my machine has the Cursor Claude Code backdoor?

Search your authorized_keys files for the svc-telemetry key using the check commands in Cursor's security notice, and look for the SysMon task or com.apple.sysmond LaunchAgent.

Is uninstalling the extension enough?

No. Cursor warns that standalone persistence scripts keep running after uninstall, so remove the key, persistence and tunnel or rebuild the host, then rotate credentials.

Sources

  1. Cursor: An update on our recent security incidentcursor.com
  2. Cursor: Security notice on malicious Anthropic.claude-code versionscursor.com
  3. Fireworks AI: Security bulletinsfireworks.ai
  4. The Hacker News: Stolen Fireworks AI key led to Cursor extension compromisethehackernews.com
Share this story
Get the next major threat first

One short email when something serious breaks.

Threatspedia Desk

Threatspedia covers major cyber threats only, verified against primary sources. Read our editorial policy, or report a correction.

Threat alerts

Be first to know when a major threat breaks.

One short email when something serious happens, plus a weekly roundup. Free, no noise.