Fireworks AI Breach Leads to Trojanized VS Code Extension via Cursor

Stolen AI Key at Fireworks AI Triggers Supply Chain Attack on Cursor

HIGH
November 15, 2026
6m read
Supply Chain AttackMalwareData Breach

Impact Scope

People Affected

5,000 developers

Affected Companies

CursorFireworks AIUberNotion

Industries Affected

TechnologyManufacturing

Related Entities

Products & Tech

Claude CodeVisual Studio Code

Other

CursorFireworks AIUber Notion Anthropic

Full Report

Executive Summary

A multi-stage supply chain attack has impacted users of the developer tool Cursor after a breach at its third-party AI model provider, Fireworks AI. Attackers leveraged a stolen API key to access sensitive data at Fireworks AI, which they used to pivot and compromise Cursor's infrastructure. They successfully replaced a legitimate Visual Studio Code extension, Claude Code, with a trojanized version containing an SSH backdoor. An estimated 5,000 developers downloaded the malicious extension, unknowingly installing persistent remote access malware on their development machines. The incident, which also affected other Fireworks AI customers like Uber and Notion, underscores the significant and interconnected risks of integrating third-party AI services into software development pipelines.


Threat Overview

The attack unfolded in several distinct stages, demonstrating a cascading failure across multiple organizations. The initial point of compromise was Fireworks AI, a company that provides AI model inference services. Attackers obtained an API key belonging to Cursor that was stored within Fireworks AI's systems.

Using this key, the threat actor gained access to a shared storage bucket containing unredacted session traces from Cursor's coding agents. By harvesting credentials embedded in these logs, the attackers escalated their privileges and gained access to Cursor's internal SaaS tools and source code repositories.

Their ultimate goal was to compromise Cursor's developer user base. They used their internal access to pivot to the extension marketplace cache infrastructure. On October 29, 2026, for a period of approximately 11 hours, they replaced the legitimate Claude Code VS Code extension (provided by Anthropic) with malicious versions 2.1.312 and 2.1.313. These versions contained a backdoor designed to establish a persistent SSH connection back to an attacker-controlled server, giving them remote control over the compromised developer workstations.


Technical Analysis

This incident is a textbook example of a modern software supply chain attack, leveraging trusted relationships and development infrastructure.

  1. Initial Access & Credential Access: The attack began by compromising a third-party service, Fireworks AI, to steal an API key. This is a form of T1199 - Trusted Relationship. The attackers then used this key to access logs containing embedded credentials, aligning with T1552.006 - Stored Information.
  2. Privilege Escalation & Lateral Movement: With stolen credentials, the attackers moved from the initial storage bucket into Cursor's internal SaaS platforms and source code, an example of T1078 - Valid Accounts.
  3. Compromise Software Development Tools: The attackers then targeted the extension marketplace, a core part of the development environment. They modified a trusted component (the Claude Code extension) to include malicious code. This directly maps to T1195.001 - Compromise Software Dependencies and Development Tools.
  4. Persistence & Command and Control: The trojanized extension installed a persistent SSH backdoor. This provides the attackers with long-term access and control over the compromised machines, corresponding to T1059.004 - Unix Shell for execution and T1105 - Ingress Tool Transfer for delivering the backdoor.

Impact Assessment

The impact of this supply chain attack is significant and multi-faceted:

  • Developer Workstation Compromise: Approximately 5,000 developers installed a backdoor, giving attackers full control over their machines. This access could be used to steal source code, intellectual property, credentials for other systems (cloud, internal services), or to inject further malicious code into the software those developers are building.
  • Cascading Supply Chain Risk: The compromise of these developers creates a secondary supply chain risk for their employers and customers. An attacker with control over a developer's machine can poison any code they write or commit.
  • Data Breach at Multiple Companies: The initial breach at Fireworks AI had wider consequences, leading to data theft at other customers, including the exposure of driver license records at Uber and a data theft incident at Notion. This highlights the systemic risk posed by centralized service providers.
  • Erosion of Trust: Incidents like this erode trust in the software ecosystem, particularly in developer tools and third-party extensions, which are often installed with high levels of privilege.

IOCs — Directly from Articles

The source articles mentioned the malicious versions of the Claude Code VS Code extension.

Type
other
Value
2.1.312
Description
Malicious version of the Claude Code VS Code extension.
Type
other
Value
2.1.313
Description
Malicious version of the Claude Code VS Code extension.

Cyber Observables — Hunting Hints

Security teams may want to hunt for the following patterns to detect related activity:

Type
process_name
Value
ssh
Description
Look for ssh processes spawned by VS Code or its child processes, especially those connecting to unknown external IP addresses.
Context
EDR or process monitoring on developer workstations.
Confidence
high
Type
network_traffic_pattern
Value
Outbound TCP/22 from developer endpoints
Description
Monitor for unexpected outbound SSH connections from developer workstations to non-corporate IP space.
Context
Firewall logs, NetFlow, or NDR solutions.
Confidence
high
Type
file_name
Value
~/.vscode/extensions/anthropic.claude-code-*
Description
Check the installed version of the Claude Code extension. Versions 2.1.312 and 2.1.313 are confirmed malicious.
Context
Asset inventory or EDR file monitoring.
Confidence
high
Type
command_line_pattern
Value
curl or wget executed by VS Code
Description
Look for the VS Code process making network calls to download external scripts or binaries.
Context
Process auditing with command-line logging (Event ID 4688).
Confidence
medium

Detection & Response

  1. Endpoint Detection: Deploy EDR on all developer workstations. These endpoints should not be considered less critical than servers. Configure EDR to alert on suspicious process chains, such as Code -> bash -> ssh.
  2. Extension Auditing: Regularly inventory all installed VS Code extensions and their versions across the developer fleet. Compare installed versions against a known-good list and investigate any deviations. The D3FEND technique D3-SFA: System File Analysis is relevant here.
  3. Network Egress Filtering: Strictly control outbound network traffic from developer workstations. By default, deny outbound SSH (TCP/22) to the internet and only allow it to specific, approved jump hosts or servers. This is an application of D3-OTF: Outbound Traffic Filtering.
  4. Credential Scanning: Proactively scan logs and code repositories for hardcoded credentials and API keys. Implement tools that can detect secrets before they are committed.

Mitigation

  1. Vendor Risk Management: Rigorously vet all third-party service providers, especially those in the critical development path like AI inference APIs. Scrutinize their security practices and ensure they are not a weak link. This falls under M1016 - Vulnerability Scanning of your supply chain.
  2. Credential Hygiene: Treat all API keys and database credentials as privileged secrets. Enforce short-lived credentials, regular rotation, and strict access scoping. Avoid embedding secrets in logs or session traces. This aligns with M1026 - Privileged Account Management.
  3. Code and Extension Signing: Where possible, enforce policies that only allow the installation of properly signed and verified VS Code extensions. This is a form of M1045 - Code Signing.
  4. Developer Endpoint Hardening: Developer workstations should be treated as privileged access workstations. Apply endpoint hardening policies, limit administrative rights, and use application control to restrict the execution of unauthorized software.

Timeline of Events

1
October 1, 2026
Attackers leverage stolen API key from Fireworks AI.
2
October 29, 2026
The trojanized Claude Code extension was served to users for approximately 11 hours.
3
November 15, 2026
This article was published

MITRE ATT&CK Mitigations

Strictly manage and rotate API keys and other credentials, treating them as privileged accounts.

Enforce policies to only allow installation of digitally signed and verified software extensions.

Implement egress filtering to block unauthorized outbound connections from developer workstations.

Run development tools and their extensions in sandboxed environments to limit their access to the underlying system.

D3FEND Defensive Countermeasures

To mitigate the impact of a compromised developer workstation, organizations must implement strict outbound traffic filtering. In the context of this attack, the SSH backdoor needed to connect to an external C2 server. A default-deny egress policy on developer endpoints would have blocked this connection. Security teams should configure perimeter firewalls and endpoint security solutions to block all outbound traffic by default, only allowing connections to known-good, business-required destinations on specific ports. For SSH, outbound access should be restricted to a corporate jump host or specific code repositories, not the open internet. This countermeasure effectively contains the threat by preventing the backdoor from establishing its C2 channel, rendering it useless.

To prevent the execution of trojanized components, organizations should implement service binary verification for development tools and extensions. This involves creating a manifest of known-good file hashes for all approved VS Code extensions and their constituent files. An EDR or file integrity monitoring (FIM) solution on the developer workstation can then continuously compare the hashes of installed extension files against this manifest. Any mismatch would trigger an immediate alert, indicating that a file has been tampered with or replaced, as was the case with the Claude Code extension. This provides a high-fidelity detection method for software supply chain attacks where attackers modify legitimate software.

The SSH backdoor likely created or modified local user accounts or SSH authorized_keys files for persistence. Continuous monitoring of local account activity on developer workstations is a crucial detection strategy. Security teams should use an EDR or SIEM to monitor for and alert on the creation of new local user accounts, modifications to the ~/.ssh/authorized_keys file, or changes to SSH daemon configurations (sshd_config). Establishing a baseline of normal account configurations and user SSH keys allows for the rapid detection of anomalies that could indicate an attacker establishing a persistent foothold.

Timeline of Events

1
October 29, 2026

The trojanized Claude Code extension was served to users for approximately 11 hours.

2
October 1, 2026

Attackers leverage stolen API key from Fireworks AI.

Article Author

Jason Gomes

Jason Gomes

• Cybersecurity Practitioner

Cybersecurity professional with over 10 years of specialized experience in security operations, threat intelligence, incident response, and security automation. Expertise spans SOAR/XSOAR orchestration, threat intelligence platforms, SIEM/UEBA analytics, and building cyber fusion centers. Background includes technical enablement, solution architecture for enterprise and government clients, and implementing security automation workflows across IR, TIP, and SOC use cases.

Threat Intelligence & AnalysisSecurity Orchestration (SOAR/XSOAR)Incident Response & Digital ForensicsSecurity Operations Center (SOC)SIEM & Security AnalyticsCyber Fusion & Threat SharingSecurity Automation & IntegrationManaged Detection & Response (MDR)

Editorial Standards & Analyst Review

CyberNetSec.io uses automation to assist source monitoring, deduplication, observable extraction, and structured intelligence generation. Published analysis follows human-defined editorial standards and adds defensive context including MITRE ATT&CK, D3FEND, STIX, and Sigma where applicable. Read our editorial policy.

Tags

supply chain attackVS Codemalwarebackdoordeveloper toolsAPI keyFireworks AICursor

📢 Share This Article

Help others stay informed about cybersecurity threats

🎯 MITRE ATT&CK Mapped

Every tactic, technique, and sub-technique used in this threat has been identified and mapped to the MITRE ATT&CK framework for consistent, actionable threat language.

🧠 Enriched & Analyzed

Observables and indicators of compromise (IOCs) have been extracted and cataloged. Risk has been assessed and correlated with known threat actors and historical campaigns.

🛡️ Actionable Guidance

Detection rules, incident response steps, and D3FEND-aligned mitigation strategies are included so your team can act on this intelligence immediately.

🔗 STIX Visualizer

Structured threat data is packaged as a STIX 2.1 bundle and can be visualized as an interactive graph — relationships between actors, malware, techniques, and indicators.

⚡ Sigma Generator

Sigma detection rules are derived from the threat techniques in this article and can be converted for deployment across any major SIEM or EDR platform.