Fireworks AI Breach Exposes Cursor Data and Spreads Malicious Claude Code Extension

cursor fireworks ai

The growing adoption of Artificial Intelligence (AI) in software development is creating new opportunities for developers and organisations. However, it is also introducing new security risks, particularly when AI platforms, third-party service providers, and development tools share sensitive information.

A recent cybersecurity incident involving Fireworks AI, Cursor, and Anthropic’s Claude Code extension highlights how a compromised third-party API key can trigger a chain of security incidents. What started with the exposure of an access key at Fireworks AI eventually allowed an attacker to access sensitive Cursor data and distribute a trojanized version of a legitimate AI coding extension.

According to Cursor, the attacker exploited stolen credentials to access internal resources and subsequently manipulated cached extension files distributed through the Cursor Marketplace. The incident raises important concerns about API key security, software supply chain attacks, third-party risk management, and the security of AI-powered development environments.

Fireworks AI Security Breach

The incident began after an attacker obtained an access key that Cursor had issued to Fireworks AI, an AI infrastructure provider.

According to Cursor’s disclosure, Fireworks AI exposed the access key following a security breach. The attacker then used this key to access a read-only storage bucket that Cursor maintained for sharing data with the service provider.

Although the storage bucket had read-only permissions, it contained sensitive information that should not have been exposed.

The compromised storage bucket included limited coding agent traces from Cursor sessions that were not operating in Privacy Mode. Most of these records dated back to 2025 and had been permitted for use in model training.

The problem was that some coding session traces contained credentials that had not been properly removed or masked before storage.

As a result, the attacker was able to obtain API credentials belonging to a small percentage of Cursor customers. Cursor stated that fewer than 0.01% of its customers had their API credentials exposed through this incident.

This highlights an important security weakness: even when an organisation uses restricted storage permissions, sensitive information stored inside that environment can still become a security risk if proper data sanitisation is not implemented.

Stolen Credentials to Access Cursor’s Internal Systems

The attackers did not stop after obtaining the exposed API credentials.

Cursor reported that the stolen credentials were used as an initial entry point to access selected Software-as-a-Service (SaaS) tools used by the company.

This allowed the attacker to access several categories of internal information, including:

  • Source code repositories: Selected archived, non-production Cursor repositories were accessed. Some of these repositories contained business contact information.

  • Customer support information: Limited customer support records and email conversations were exposed.

  • Product usage information: Certain product usage metrics and customer contact details were accessed.

  • AI conversation summaries: Summaries of some coding agent conversations were obtained.

  • Customer-submitted credentials: Some credentials that users had previously shared with Cursor’s support team were also exposed.

This part of the incident demonstrates how compromised credentials can become a gateway to multiple internal systems.

In modern cloud environments, organisations often depend on several external SaaS platforms for development, collaboration, customer support, and infrastructure management. If an attacker obtains valid credentials associated with these services, the attacker may be able to move beyond the original compromised environment.

This is particularly concerning when organisations do not maintain strong identity controls, access restrictions, and continuous monitoring across their cloud and SaaS ecosystems.

Trojanized Claude Code Extension Distributed Through Cursor Marketplace

One of the most serious developments occurred on October 29, 2026.

According to Cursor, the attacker used credentials obtained from the earlier compromise to place malicious versions of Anthropic’s official Claude Code extension into the cache used to serve the Cursor extension marketplace.

The malicious extension was designed to appear like a legitimate software update.

However, its actual purpose was to install a persistent SSH backdoor on affected systems.

An SSH backdoor can provide attackers with an additional method of accessing a compromised machine. Depending on its implementation and the privileges available, it may allow unauthorised remote access and provide an opportunity for further malicious activities.

The incident is particularly concerning because developers generally trust extensions distributed through official marketplaces. Such extensions often receive access to source code, development environments, terminals, and other resources required for their functionality.

A compromised extension can therefore become an effective attack vector against developer workstations and potentially sensitive development environments.

Cursor stated that up to 5,000 users who requested specific cached versions of the extension on October 29 could potentially have received the malicious build.

The malicious versions remained available for approximately 11 hours before Cursor removed them.

The affected extension versions were identified as:

  • Version 2.1.312

  • Version 2.1.313

Cursor subsequently took the extension marketplace offline and worked with the cloud service provider hosting the attacker’s server to take down the malicious infrastructure.

The company also notified affected users who had installed the identified versions during the exposure period and assisted them in revoking potentially exposed credentials.

Cursor’s Response and Security Measures

Following the discovery, Cursor initiated several containment and remediation activities to limit the impact of the incident.

The company deactivated the compromised Fireworks AI access key and revoked and rotated relevant credentials. It also investigated potential unauthorised access to other internal systems.

Cursor temporarily disabled certain internal and product services during the investigation, which resulted in disruptions to some platform features.

The company also took additional steps to reduce its dependency on the compromised service provider.

These included:

  • Moving all AI inference traffic away from Fireworks AI.

  • Deactivating the access key issued to Fireworks AI.

  • Revoking its access to datasets previously shared with the provider.

  • Investigating potential deeper access to internal systems.

  • Removing the malicious extension builds and taking the marketplace temporarily offline.

Cursor further stated that it had not detected additional malicious activity since November 1, 2026.

The company also clarified that there was no evidence of impact to users who had enabled Privacy Mode, customer code repositories, its production environment, the Cursor application and official updates, or other extensions and plugins downloaded from the marketplace.

artificial intelligence ai

The Fireworks AI breach appears to have affected more than one organisation.

Uber confirmed that an unauthorised third party accessed its account with a service provider after compromising the provider’s internal credentials.

According to a filing submitted to the California authorities, Uber was notified of the service provider compromise on October 9, 2026. Its investigation subsequently determined that the attacker had accessed Uber-related data through the compromised service provider account.

Uber also confirmed that thousands of drivers’ licence records were stolen.

Although Uber did not publicly identify the service provider in its disclosure, reporting by TechCrunch linked the incident to Fireworks AI.

Another organisation, Notion, also confirmed that its data had been stolen in the Fireworks AI incident.

These developments demonstrate how a single third-party security failure can potentially affect multiple organisations that depend on the same service provider.

This is a classic example of a third-party supply chain security risk, where an organisation’s security posture is influenced not only by its own controls but also by the security practices of its technology partners.

Key Lessons

The Cursor and Fireworks AI incident provides several important lessons for organisations adopting AI-powered development platforms and cloud-based services.

1. Strengthen API Key and Secret Management

API keys, access tokens, and service credentials should never be stored in plain text or exposed through application logs, development traces, or training datasets. Organisations should implement centralised secrets management, automatic credential rotation, and continuous secret scanning.

2. Implement Strong Third-Party Risk Management

Vendor security assessments should go beyond annual questionnaires and compliance certifications. Organisations must understand what data is shared with third parties, how it is protected, who can access it, and how quickly access can be revoked following a security incident.

3. Apply Data Sanitisation Before AI Model Training

AI-powered applications frequently process source code, prompts, development logs, and conversation histories. Sensitive information such as API keys, passwords, access tokens, and personal data must be identified and removed before such information is stored or used for model training.

4. Secure Software Supply Chains

Software extension marketplaces and development plugins should be treated as potential attack surfaces. Organisations should implement extension allowlisting, version verification, software provenance checks, and endpoint monitoring to identify suspicious modifications.

5. Adopt Zero Trust Access Controls

Access to internal SaaS platforms, repositories, and cloud resources should be restricted according to the principle of least privilege. Multi-factor authentication, conditional access, short-lived credentials, and continuous access monitoring can reduce the impact of compromised identities.

6. Improve Monitoring of Developer Environments

Security teams should monitor unusual SSH activity, unexpected extension installations, suspicious credential usage, and unauthorised access to development resources. Developer workstations must receive the same security attention as other critical enterprise endpoints.

The Fireworks AI and Cursor incident is another reminder that cybersecurity risks are evolving alongside the rapid adoption of artificial intelligence.

The attack did not begin with a sophisticated vulnerability in an AI model. Instead, it started with an exposed access key at a third-party service provider and progressed through compromised credentials, internal SaaS access, and the distribution of a malicious software extension.

For organisations, the bigger lesson is that AI security cannot be limited to protecting AI models and applications. It must also cover the entire ecosystem of identities, APIs, datasets, development tools, cloud platforms, and third-party service providers.

As organisations increasingly integrate AI into software development and business operations, securing these interconnected environments will become an essential part of enterprise cybersecurity strategy.

Follow us on Twitter and Linkedin for real time updates and exclusive content.

Scroll to Top