On November 21, 2026, Cursor confirmed that the official Claude Code extension distributed through its in-app extension marketplace was trojanized after an attacker leveraged a stolen Fireworks AI access key. The key — originally issued by Cursor to Fireworks AI as part of a backend integration — was exposed during a separate breach targeting Fireworks AI. The attacker then used the key's read-only access to copy data and, critically, to push a malicious version of the Claude Code extension to Cursor's marketplace, where it was served to users as a legitimate update.
This incident is a textbook supply-chain compromise: a trusted vendor integration became the attack vector, and the payload was delivered through a channel developers implicitly trust — their IDE's extension marketplace. Any organization whose developers installed or updated the Claude Code extension during the compromise window must assume exposure of source code, credentials, environment variables, and potentially production secrets accessible from developer workstations.
Severity: Critical for development environments. Developer workstations are high-value targets with access to source repositories, CI/CD tokens, cloud credentials, and internal networks. A trojanized extension running in the IDE context inherits all of those permissions.
Technical Analysis
Attack Chain
- Initial Compromise (Fireworks AI): Fireworks AI, an AI inference platform, suffered a breach. During that incident, an access key that Cursor had issued to Fireworks was exposed to the attacker.
- Key Abuse: The attacker used the stolen key to authenticate to Cursor's systems. Cursor described the key as "read-only" in their Friday update, but the attacker was able to "copy data" and subsequently interact with the extension publishing pipeline.
- Extension Trojanization: The attacker pushed a modified version of the Claude Code extension to Cursor's extension marketplace. The trojanized extension was served to users as a routine update from a trusted publisher (Anthropic's official extension).
- Execution on Developer Machines: Developers running Cursor who had the Claude Code extension installed would have automatically or manually updated to the malicious version. The extension executes with the full privileges of the Cursor IDE process — which on a developer workstation typically includes access to SSH keys, cloud CLI credentials,
.envfiles, source code, and internal network resources.
Affected Products
- Cursor IDE (all platforms: Windows, macOS, Linux)
- Claude Code extension distributed via Cursor's extension marketplace (versions published during the compromise window)
- Fireworks AI integration with Cursor (the stolen access key's scope)
Exploitation Status
- Confirmed active exploitation. The trojanized extension was distributed through Cursor's official marketplace to end users.
- No CVE has been assigned — this is a credential theft and supply-chain abuse scenario, not a software vulnerability in the traditional sense. The root cause is a leaked access key and insufficient scoping/segmentation of that key's permissions.
- Not currently listed in CISA KEV (as of this writing), but the incident pattern (supply-chain extension compromise) mirrors prior high-impact events.
Why This Matters for Defenders
This is not just a Cursor problem. The underlying failure — an inter-service access key exposed in a third-party breach, then used to compromise a distribution pipeline — is endemic to modern SaaS and AI toolchain integrations. Any organization that issues API keys, service principals, or OAuth tokens to third-party vendors faces the same risk profile.
Detection & Response
The primary detection surface is developer endpoints. The trojanized extension runs within the Cursor process, but a malicious payload will likely spawn child processes, make outbound network connections, and access credential stores. Focus on anomalous process trees, unexpected network destinations from IDE processes, and extension file integrity.
SIGMA Rules
---
title: Cursor IDE Spawning Suspicious Child Processes
id: 8b2e4f1a-3c7d-4e9a-b5f6-2a1c8d4e7f90
status: experimental
description: Detects Cursor IDE or its extension host spawning shell interpreters, script engines, or download tools — consistent with a trojanized extension executing payloads.
references:
- https://thehackernews.com/2026/11/stolen-fireworks-ai-key-led-to-cursor.html
- https://attack.mitre.org/techniques/T1059/
author: Security Arsenal
date: 2026/11/22
tags:
- attack.execution
- attack.t1059
- attack.supply_chain_compromise
- attack.t1195.002
logsource:
category: process_creation
product: windows
detection:
selection_parent:
ParentImage|endswith:
- '\cursor.exe'
- '\Cursor.exe'
- '\Cursor Helper.exe'
selection_child:
Image|endswith:
- '\cmd.exe'
- '\powershell.exe'
- '\pwsh.exe'
- '\wscript.exe'
- '\cscript.exe'
- '\curl.exe'
- '\certutil.exe'
- '\bitsadmin.exe'
condition: selection_parent and selection_child
falsepositives:
- Legitimate Claude Code extension functionality invoking terminal commands — tune by CommandLine content
level: high
---
title: Cursor IDE Accessing Credential Stores
id: 3f9a1c5e-7b2d-4e8f-a6c3-1d5e9b2f4a78
status: experimental
description: Detects Cursor IDE processes accessing browser credential stores, SSH keys, or cloud credential files — indicative of a trojanized extension harvesting secrets.
references:
- https://thehackernews.com/2026/11/stolen-fireworks-ai-key-led-to-cursor.html
- https://attack.mitre.org/techniques/T1552/
author: Security Arsenal
date: 2026/11/22
tags:
- attack.credential_access
- attack.t1552.001
- attack.t1552.004
logsource:
category: file_access
product: windows
detection:
selection_process:
Image|endswith:
- '\cursor.exe'
- '\Cursor.exe'
- '\Cursor Helper.exe'
selection_target:
TargetFilename|contains:
- '\.ssh\'
- '\.aws\credentials'
- '\.azure\'
- '\.config\gcloud\'
- '\Login Data'
- '\Cookies'
- '\.env'
- '\.npmrc'
- '\.netrc'
- '\kube\config'
condition: selection_process and selection_target
falsepositives:
- Legitimate extension functionality accessing .env files or SSH config for dev workflows — establish baseline before alerting
level: high
---
title: Cursor Extension Directory Modification
id: 6c4d8e2a-9f1b-4d5e-a3c7-8b6f2d1e4a90
status: experimental
description: Detects file modifications within Cursor extension directories by non-Cursor processes — may indicate extension replacement or tampering outside the normal update mechanism.
references:
- https://thehackernews.com/2026/11/stolen-fireworks-ai-key-led-to-cursor.html
- https://attack.mitre.org/techniques/T1195/
author: Security Arsenal
date: 2026/11/22
tags:
- attack.supply_chain_compromise
- attack.t1195.002
- attack.persistence
logsource:
category: file_event
product: windows
detection:
selection_path:
TargetFilename|contains:
- '\.cursor\extensions\'
- '\Cursor\User\globalStorage\'
exclusion_updater:
Image|endswith:
- '\cursor.exe'
- '\Cursor.exe'
- '\Cursor Helper.exe'
- '\Code.exe'
condition: selection_path and not exclusion_updater
falsepositives:
- Manual extension installation by developers
- Endpoint management software
level: medium
KQL (Microsoft Sentinel / Defender)
// Hunt for Cursor IDE spawning suspicious child processes across all endpoints
// This query looks for shell interpreters, script engines, and download tools
// spawned by Cursor or its extension host — consistent with trojanized extension behavior.
// Requires: Defender for Endpoint (DeviceProcessEvents)
let Lookback = 14d;
let CursorProcesses = dynamic(["cursor.exe", "Cursor.exe", "Cursor Helper.exe", "Cursor Helper (Renderer).exe"]);
let SuspiciousChildren = dynamic(["cmd.exe", "powershell.exe", "pwsh.exe", "wscript.exe", "cscript.exe", "curl.exe", "certutil.exe", "bitsadmin.exe", "mshta.exe", "rundll32.exe"]);
DeviceProcessEvents
| where TimeGenerated > ago(Lookback)
| where InitiatingProcessFileName in~ (CursorProcesses)
| where FileName in~ (SuspiciousChildren)
| project TimeGenerated, DeviceName, InitiatingProcessFileName, InitiatingProcessCommandLine,
FileName, ProcessCommandLine, AccountName, FolderPath, SHA256
| extend IsDownloadTool = iff(FileName in~ ("curl.exe", "certutil.exe", "bitsadmin.exe"), true, false)
| sort by TimeGenerated desc;
// Hunt for outbound network connections from Cursor to non-standard destinations
// Exclude known-good Cursor/Anthropic API endpoints; alert on anything else
DeviceNetworkEvents
| where TimeGenerated > ago(Lookback)
| where InitiatingProcessFileName in~ (CursorProcesses)
| where RemoteUrl !has_any ("anthropic.com", "cursor.sh", "cursor.com", "fireworks.ai", "microsoft.com")
| where RemotePort in (443, 8443, 4444, 8080, 1337, 6667)
| summarize Connections = count(), DistinctDestinations = dcount(RemoteIP)
by DeviceName, RemoteIP, RemoteUrl, RemotePort, InitiatingProcessFileName
| where Connections > 5
| sort by DistinctDestinations desc;
Velociraptor VQL
-- Hunt for trojanized Cursor extension artifacts and suspicious child processes
-- Run this artifact across all endpoints with Cursor installed
-- Section 1: Enumerate installed Cursor extensions and check for recent modifications
SELECT FullPath, Size, Mtime, Atime, Ctime,
hash(path=FullPath).SHA256 AS FileHash
FROM glob(globs="C:\\Users\\*\\.cursor\\extensions\\**\\package.json")
WHERE Mtime > ago(30 * 24 * 60 * 60)
ORDER BY Mtime DESC
-- Section 2: Check for suspicious processes spawned by Cursor
SELECT Pid, Ppid, Name, CommandLine, Exe, Username, CreateTime
FROM pslist()
WHERE Ppid IN (
SELECT Pid FROM pslist() WHERE Name =~ '(?i)cursor'
)
AND Name =~ '(?i)(cmd|powershell|pwsh|wscript|cscript|curl|certutil|mshta|rundll32)'
-- Section 3: Check Cursor extension host network connections
SELECT Pid, Name, Status, LocalAddr, LocalPort, RemoteAddr, RemotePort
FROM netstat()
WHERE Name =~ '(?i)cursor'
AND Status = 'ESTABLISHED'
AND RemoteAddr !~ '^(10\\.|172\\.(1[6-9]|2[0-9]|3[01])\\.|192\\.168\\.|127\\.)'
Remediation Script
#Requires -RunAsAdministrator
<#
.SYNOPSIS
Cursor Claude Code Extension Compromise — Endpoint Audit and Remediation Script
.DESCRIPTION
Audits developer workstations for the trojanized Claude Code extension,
checks for suspicious child processes, and remediates by removing the
compromised extension and rotating exposed credentials.
.NOTES
Security Arsenal — November 2026
Reference: https://thehackernews.com/2026/11/stolen-fireworks-ai-key-led-to-cursor.html
#>
$ErrorActionPreference = 'Continue'
$ReportPath = "$env:SystemDrive\Cursor_Extension_Audit_$(Get-Date -Format 'yyyyMMdd_HHmmss').log"
function Write-Finding {
param([string]$Message, [string]$Severity = 'INFO')
$entry = "[$(Get-Date -Format 'yyyy-MM-dd HH:mm:ss')] [$Severity] $Message"
Add-Content -Path $ReportPath -Value $entry
Write-Host $entry
}
Write-Finding "=== Cursor Extension Compromise Audit Starting ===" "INFO"
# --- Step 1: Check if Cursor is installed ---
$cursorPaths = @(
"$env:LOCALAPPDATA\Programs\cursor\Cursor.exe",
"$env:ProgramFiles\Cursor\Cursor.exe"
)
$cursorInstalled = $false
foreach ($path in $cursorPaths) {
if (Test-Path $path) {
$cursorInstalled = $true
$cursorVersion = (Get-Item $path).VersionInfo.ProductVersion
Write-Finding "Cursor found: $path (Version: $cursorVersion)" "INFO"
}
}
if (-not $cursorInstalled) {
Write-Finding "Cursor IDE not found on this system. No action needed." "INFO"
exit 0
}
# --- Step 2: Enumerate installed extensions ---
$extensionDirs = @(
"$env:USERPROFILE\.cursor\extensions",
"$env:APPDATA\Cursor\User\globalStorage"
)
foreach ($dir in $extensionDirs) {
if (Test-Path $dir) {
Write-Finding "Scanning extension directory: $dir" "INFO"
$extensions = Get-ChildItem -Path $dir -Directory -ErrorAction SilentlyContinue
foreach ($ext in $extensions) {
$lastWrite = $ext.LastWriteTime
Write-Finding " Extension: $($ext.Name) | Last Modified: $lastWrite" "INFO"
# Flag Claude Code extension specifically
if ($ext.Name -match '(?i)claude|anthropic') {
Write-Finding " *** ALERT: Claude/Anthropic extension found: $($ext.Name) ***" "CRITICAL"
Write-Finding " Last modified: $lastWrite — verify against Cursor's published compromise window" "CRITICAL"
# Check for suspicious files within extension
$suspiciousFiles = Get-ChildItem -Path $ext.FullName -Recurse -Include *.exe,*.ps1,*.bat,*.cmd,*.vbs,*.js -ErrorAction SilentlyContinue |
Where-Object { $_.LastWriteTime -gt (Get-Date).AddDays(-30) }
foreach ($file in $suspiciousFiles) {
$hash = (Get-FileHash -Path $file.FullName -Algorithm SHA256).Hash
Write-Finding " Recent file: $($file.Name) | SHA256: $hash" "WARNING"
}
}
}
}
}
# --- Step 3: Check for suspicious child processes of Cursor ---
Write-Finding "Checking for suspicious child processes of Cursor..." "INFO"
$suspiciousChildren = @('cmd', 'powershell', 'pwsh', 'wscript', 'cscript', 'curl', 'certutil', 'mshta', 'rundll32', 'bitsadmin')
$cursorPids = Get-Process -Name 'Cursor*' -ErrorAction SilentlyContinue | Select-Object -ExpandProperty Id
foreach ($pid_ in $cursorPids) {
$children = Get-CimInstance Win32_Process -Filter "ParentProcessId = $pid_" -ErrorAction SilentlyContinue
foreach ($child in $children) {
$childName = $child.Name -replace '\.exe$',''
if ($childName -in $suspiciousChildren) {
Write-Finding " SUSPICIOUS: Cursor (PID $pid_) spawned $($child.Name) (PID $($child.ProcessId))" "CRITICAL"
Write-Finding " CommandLine: $($child.CommandLine)" "CRITICAL"
}
}
}
# --- Step 4: Check network connections from Cursor ---
Write-Finding "Checking active network connections from Cursor processes..." "INFO"
foreach ($pid_ in $cursorPids) {
$connections = Get-NetTCPConnection -OwningProcess $pid_ -State Established -ErrorAction SilentlyContinue |
Where-Object { $_.RemoteAddress -notmatch '^(10\.|172\.(1[6-9]|2[0-9]|3[01])\.|192\.168\.|127\.|::1|fe80)' }
foreach ($conn in $connections) {
Write-Finding " External connection: $($conn.RemoteAddress):$($conn.RemotePort) (PID $pid_)" "WARNING"
}
}
# --- Step 5: Remediation prompt ---
Write-Finding "" "INFO"
Write-Finding "=== REMEDIATION RECOMMENDATIONS ===" "INFO"
Write-Finding "1. If Claude Code extension is installed and was updated during the compromise window:" "ACTION"
Write-Finding " a. Remove the extension immediately: rm -rf ~/.cursor/extensions/*claude*" "ACTION"
Write-Finding " b. Update Cursor to the latest version (check cursor.com for patched release)" "ACTION"
Write-Finding " c. Reinstall Claude Code extension ONLY from verified source after Cursor confirms fix" "ACTION"
Write-Finding "2. Rotate ALL credentials accessible from this workstation:" "ACTION"
Write-Finding " - SSH keys (~/.ssh/)" "ACTION"
Write-Finding " - Cloud credentials (~/.aws/credentials, ~/.azure/, gcloud)" "ACTION"
Write-Finding " - API tokens in .env files" "ACTION"
Write-Finding " - Git credentials and personal access tokens" "ACTION"
Write-Finding " - Browser-stored passwords if accessible" "ACTION"
Write-Finding "3. Check source code repositories for unauthorized commits or access" "ACTION"
Write-Finding "4. Review CI/CD pipeline logs for anomalous activity" "ACTION"
Write-Finding "=== Audit Complete. Report saved to: $ReportPath ===" "INFO"
Remediation
Immediate Actions (24-48 Hours)
- Identify exposed workstations: Inventory all developer machines with Cursor IDE installed. Determine which had the Claude Code extension active during the compromise window.
- Remove the compromised extension: Uninstall the Claude Code extension from all affected machines. Do not simply disable — remove the extension files entirely.
- Update Cursor: Cursor has presumably pushed a clean version. Update to the latest release and verify the extension publisher and signature before reinstalling.
- Rotate all developer credentials: Treat every secret accessible from affected workstations as compromised:
- SSH private keys (
~/.ssh/) - Cloud CLI credentials (AWS, Azure, GCP)
- API tokens and keys in
.envfiles - Git credentials and personal access tokens
- Kubernetes configs (
~/.kube/config) - npm/pip registry tokens
- SSH private keys (
- Audit source repositories: Check for unauthorized commits, new branches, or modified CI/CD configurations.
Medium-Term Actions (1-2 Weeks)
- Review third-party API key scope: Cursor's key was described as "read-only" yet the attacker pushed a malicious extension. Audit your own organization's service-to-service keys and ensure least-privilege scoping. Read-only keys should never have write access to distribution pipelines.
- Implement extension allowlisting: If your organization uses Cursor or VS Code-based IDEs, consider enforcing an extension allowlist via group policy or MDM to prevent unauthorized extension installations.
- Deploy extension integrity monitoring: Use file integrity monitoring (FIM) on extension directories to detect unauthorized modifications.
- Rotate Fireworks AI keys: If your organization uses Fireworks AI, verify whether your keys were exposed in their breach and rotate accordingly.
Long-Term Actions
- Vendor risk assessment: Evaluate the security posture of AI/ML toolchain vendors (Cursor, Fireworks, Anthropic, etc.) with the same rigor as traditional software suppliers.
- Developer endpoint hardening: Developer workstations should be treated as Tier-0 assets. Implement EDR with process lineage tracking, restrict outbound network access to known-good destinations, and enforce secrets management (vault-based, not file-based).
- Incident response plan update: Add "trojanized IDE extension" as a scenario in your IR playbooks.
Key Takeaways
- Third-party key exposure is a first-party problem. Cursor issued a key to Fireworks AI. When Fireworks was breached, Cursor's distribution pipeline was compromised. Every API key you issue to a vendor is an extension of your attack surface.
- "Read-only" is not always read-only. Cursor described the key as read-only, yet the attacker pushed a malicious extension. Verify the actual API permissions and pipeline access your service keys grant — do not trust labels.
- IDE extensions are supply-chain vectors. Developer tools have deep access to credentials, source code, and internal networks. A compromised extension is functionally equivalent to a compromised developer.
- Credential rotation is non-negotiable. If a developer workstation was running the trojanized extension, every secret on that machine must be rotated — no exceptions.
Related Resources
Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.