Cursor Cuts Ties With Fireworks AI After Security Breach Exposes Customer Data
Share
Cursor has ended its reliance on Fireworks AI for model inference after an exposed vendor access key allowed attackers to access historical AI coding-agent data, steal credentials, and compromise parts of its internal systems. The attackers later used stolen credentials to distribute malicious versions of Anthropic’s Claude Code extension through Cursor’s VS Code extension marketplace.
In a security update published on November 13, 2026, Cursor explained how the attack began with an access key it had issued to Fireworks AI, an external provider supporting its AI infrastructure. The exposed key gave attackers access to a storage bucket containing historical coding-agent traces. Some of these records included API credentials that hadn’t been properly redacted.
The attackers used the exposed credentials to access selected software-as-a-service (SaaS) tools used by Cursor. They eventually obtained access to internal data and used another compromised credential to distribute malicious extensions to developers.
Cursor estimates that up to 5,000 users who requested the affected extension versions during the exposure window may have received the malicious builds.
The incident highlights the risks of relying on third-party AI infrastructure providers, particularly when those providers handle sensitive customer data and credentials.
How the Attack Started
The attack began when cybercriminals obtained an access key that Cursor had issued to Fireworks AI. According to Cursor’s investigation, Fireworks exposed the key, allowing attackers to access a read-only storage bucket that Cursor used to share data with the provider.
The bucket contained historical AI coding-agent traces, mostly from 2025. These records came from sessions where users hadn’t enabled Privacy Mode and where Cursor was permitted to use the data for model training.
Some of the traces contained credentials that hadn’t been properly redacted. The attackers extracted these credentials and used them to access selected SaaS tools connected to Cursor’s internal operations.
The compromised information included:
- Archived, non-production Cursor source code repositories, some containing business contact information.
- Limited customer support records and email conversations.
- Product usage metrics and customer contact details.
- Summaries of certain AI coding-agent conversations.
- Credentials that customers had previously shared with Cursor’s support team.
Cursor reported that fewer than 0.05% of users had their coding-agent traces exposed. Less than 0.01% of customers had API credentials compromised.
Malicious Claude Code Extensions Distributed Through Cursor
One of the most serious developments involved the distribution of unauthorized Claude Code extensions through Cursor’s VS Code extension marketplace.
On October 29, attackers used a credential obtained from the compromised internal data to place malicious copies of Anthropic’s Claude Code extension in the cache serving Cursor’s marketplace.
The affected versions were 2.1.312 and 2.1.313. The attackers modified the legitimate extension by inserting malicious JavaScript designed to install persistent SSH backdoors on developers’ computers.
The malware attempted to:
- Collect information about infected systems.
- Install an attacker-controlled SSH key.
- Enable SSH access where possible.
- Establish persistence mechanisms to maintain remote access.
The malicious extensions remained available for approximately 11 hours before Cursor removed them. The company estimates that up to 5,000 users who requested the affected versions during that period may have received the compromised builds.
Cursor clarified that Anthropic didn’t publish these unauthorized versions. The incident also didn’t involve a vulnerability in Anthropic’s legitimate Claude Code product. The Open VSX marketplace, which serves as the source for Cursor’s extensions, wasn’t compromised either.
What Information Was Exposed?
The attackers accessed more than historical coding-agent traces. They also reached selected internal systems and obtained information stored in Cursor’s connected services. The exposed material included limited customer support information, email conversations, product usage metrics, business contact details, and credentials shared with Cursor’s support team. However, Cursor clarified that the attackers didn’t access its customer code repositories or production infrastructure.
The company also confirmed that active inference requests weren’t exposed. The compromised coding-agent traces were limited to historical data stored for offline analysis. Users who had enabled Privacy Mode weren’t affected by the exposure of coding-agent traces described in the incident report.
How Cursor Responded to the Breach
Cursor says its monitoring systems detected unauthorized activity. The company immediately deactivated the exposed Fireworks access key and began investigating how far the attackers had moved within its systems. It also temporarily disabled certain internal and product services as a precaution while its security team investigated the incident.
Cursor has since taken several steps to contain the attack:
- Ended its reliance on Fireworks: Cursor moved all AI inference traffic away from Fireworks and revoked the provider’s access to shared datasets.
- Rotated credentials: The company revoked exposed credentials and rotated credentials across its internal systems as a precaution.
- Reviewed the extension marketplace: Cursor validated other extensions and rebuilt its VS Code extension marketplace.
- Strengthened access restrictions: The company expanded network restrictions for service providers to prevent compromised credentials from being used outside their intended environments.
- Introduced additional secret scanning: Cursor expanded secret-scanning protections across its code repositories and other internal data sources.
- Redeployed production infrastructure: The company redeployed its production infrastructure and reviewed internal services before restoring them.
- Brought in external investigators: Cursor engaged external forensic experts to support its investigation and contacted affected customers.
Cursor believes the incident is contained. The company reported that the last attacker activity it observed occurred on November 1 and was blocked.
What Cursor Users Should Do
Cursor has contacted customers whose information may have been exposed. However, developers who installed or updated Claude Code through Cursor’s marketplace on October 29 should check whether they received either of the affected versions.
The compromised versions were 2.1.312 and 2.1.313.
Uninstalling the extension alone may not remove the threat. Users who installed these versions should follow Cursor’s official remediation instructions and check their systems for suspicious SSH keys, unauthorized remote access configurations, and other signs of persistence.
Developers should also rotate potentially compromised API keys and credentials, particularly those previously shared with Cursor’s support team. They should review connected SaaS accounts for suspicious activity and revoke credentials that may have been exposed.
Cursor has published remediation guidance and scripts to help affected users investigate their systems.
The Fireworks AI Connection: Notion Faced a Similar Data Exposure
The Cursor incident isn’t an isolated case of data exposure involving Fireworks AI. Notion recently disclosed that a separate security incident at the same AI infrastructure provider exposed customer AI chat transcripts.
The exposed information included conversations processed through Notion’s Insights feature, which helps workspace administrators understand how employees use AI. Notion confirmed that fewer than 1% of its workspaces were affected.
The incident also raised questions about Fireworks AI’s data protection practices. Notion had a contractual zero-data-retention commitment from the provider, yet attackers accessed a dataset containing customer conversations.
Read our coverage of the Notion AI data breach that exposed customer chat transcripts.
Both incidents point to a common security concern. Businesses increasingly depend on external AI infrastructure providers to process customer information, internal conversations, and developer activity. A security failure at one provider can expose information belonging to several companies that rely on its services.
The Cursor incident adds another dimension to this problem. Attackers didn’t stop at accessing stored data. They used exposed credentials to reach additional systems and distribute malicious software through a trusted development tool.
Why This Incident Matters to AI Security
AI-powered development tools often handle information that attackers can use to compromise business systems. Coding-agent conversations may contain API keys, authentication tokens, internal code references, and details about application infrastructure.
When companies share this information with external AI providers, they introduce additional access points that require proper security controls.
The Cursor breach highlights three important concerns:
- Third-party access needs tighter controls
AI providers should receive only the access required to perform their assigned tasks. Companies should restrict vendor credentials, monitor their usage, and revoke access when it is no longer needed.
- Coding-agent traces require careful handling
AI conversations and development logs can contain sensitive information. Companies must implement reliable secret-scanning and redaction processes before storing or sharing these records with external providers.
- Trusted development tools can become attack channels
The distribution of malicious Claude Code extensions shows how attackers can exploit trust in software marketplaces. Developers often install extensions to improve productivity, but compromised marketplace credentials can turn these trusted tools into malware delivery channels.
The incident also reinforces the need for organizations to review their AI vendors’ security practices rather than relying solely on contractual assurances.
To Sum Up
Cursor’s decision to move its inference traffic away from Fireworks AI follows a security incident that exposed historical coding-agent data, internal information, and customer credentials. The attackers used the compromised credentials to distribute malicious Claude Code extensions, potentially exposing thousands of developers to SSH backdoors.
The incident shows how a security failure at an external AI provider can affect the companies that depend on its services. For businesses adopting AI coding assistants, protecting source code and production systems isn’t enough. They must also secure the data shared with AI providers, control third-party access, and monitor the software supply chain that supports their development tools.
As AI adoption grows, vendor security must become an essential part of every organization’s cybersecurity strategy.
