Cursor Security Breach Exposes Developer Credentials and Spreads Malicious Claude Code Extensions in Supply Chain Attack

Popular AI-powered code editor Cursor has formally confirmed that it fell victim to a sophisticated security breach. The incident enabled malicious actors to infiltrate internal company systems, exfiltrate sensitive customer data, and distribute a compromised version of Anthropic’s Claude Code extension through the official marketplace. The intrusion originated from a compromised access key stored by a third-party AI infrastructure provider, setting off a cascade of security failures that ultimately exposed thousands of software developers to persistent malware.
The breach underscores a growing and systemic vulnerability within the modern software supply chain, particularly regarding the third-party integrations and AI development platforms that have become foundational to contemporary software engineering. According to a detailed incident report published by Cursor, the compromise impacted a small fraction of its user base—affecting fewer than 0.05% of users for historical coding-agent traces and fewer than 0.01% for API credentials. However, the vector of attack and the potential consequences for developers highlight the severe risks associated with compromised AI toolchains.
The Anatomy of the Breach: From Third-Party Infrastructure to Internal Systems
The security event began when unauthorized actors gained access to an authentication key maintained on Fireworks AI, an external artificial intelligence infrastructure provider utilized by Cursor for model training and inference workloads. The exposed access key granted the hackers unauthorized entry into a read-only cloud storage bucket. Cursor utilized this bucket to share unstructured datasets with Fireworks AI for training purposes.
Crucially, the bucket contained historical AI coding-agent traces predominantly gathered throughout 2025. These logs originated from users who had not explicitly enabled "Privacy Mode," meaning their interactions were logged to help improve machine learning models. Within these historical traces, security auditors discovered
unredacted API credentials and sensitive tokens. Equipped with these keys, the attackers bypassed initial perimeters and moved laterally into several of Cursor’s internal software-as-a-service (SaaS) tools and administrative systems.
While Cursor’s internal monitoring systems quickly flagged the anomalous behavior—prompting engineers to immediately revoke the compromised Fireworks access key, rotate all exposed authentication secrets, and temporarily disable various internal and product services as a precautionary measure—the damage had already been partially realized. The attackers had successfully acquired a high-privilege credential associated with the company’s extension marketplace infrastructure.
The Supply Chain Infiltration: Malicious Claude Code Extensions
Armed with the stolen marketplace credential, the threat actors executed the most visible phase of the attack on October 29. Without authorization, they published manipulated copies of Anthropic’s legitimate Claude Code extension directly to the cache server utilized by the Cursor Visual Studio Code (VS Code) extension marketplace.

The malicious payloads were categorized under versions 2.1.312 and 2.1.313 of the Claude Code extension. They remained accessible to unsuspecting developers for approximately 11 hours before Cursor security personnel detected the anomaly and purged the files from the marketplace. Despite the relatively short window of exposure, Cursor estimates that up to 5,000 developers who requested updates or installed the extension during this timeframe may have downloaded the compromised builds.
The attack did not compromise Anthropic’s official, legitimate extension repositories, nor did it impact the independent Open VSX registry, demonstrating that the intrusion was strictly isolated to Cursor’s specific caching and distribution pipeline.
Technical Analysis of the Malware Payload
Security researchers and subsequent technical disclosures revealed that the malicious modifications made to the Claude Code extension were designed to establish a persistent foothold on developer workstations. When executed, the modified extension ran a malicious JavaScript payload engineered to harvest system information and deploy persistent Secure Shell (SSH) backdoors.
The script attempted to install an attacker-controlled SSH key onto the victim’s machine, reconfigure local network or system settings to facilitate unauthorized remote SSH access, and establish persistence mechanisms intended to survive system reboots. However, independent security analysis later uncovered fundamental coding errors within the malware script. These software defects likely thwarted the execution of the backdoor on a significant number of Windows and macOS systems, particularly those where the SSH service was not already pre-enabled and running.
Cursor emphasized that despite the deployment of the malicious extension, core development assets remained secure. The company’s primary production code infrastructure, official application binary releases, customer source code repositories, and all prompts or code belonging to users who had activated Privacy Mode were verified as completely unaffected by the incident.
Chronology of the Incident and Remediation Steps
The unfolding of the Cursor security breach followed a rapid sequence of detection, containment, and systemic rebuilding:
- Throughout 2025: Historical AI coding-agent traces, including inadvertent plaintext API keys, are logged into a read-only storage bucket shared with third-party infrastructure.
- Mid-October 2025: Threat actors compromise Fireworks AI, locate the exposed access key, and exfiltrate data from Cursor’s shared storage bucket.
- October 29 (Hours leading to the incident): Using harvested SaaS and administrative credentials, attackers acquire authorization tokens for the Cursor VS Code marketplace cache.
- October 29 (Marketplace Injection): Malicious versions (2.1.312 and 2.1.313) of the Anthropic Claude Code extension are published to the marketplace cache.
- October 29 (Detection and Removal): Cursor monitoring systems detect unauthorized access; marketplace assets are scrubbed, removing the malicious extensions after roughly 11 hours of availability.
- Post-Incident: Cursor halts inference traffic routed through Fireworks AI, invalidates and rotates all exposed credentials, rebuilds the marketplace architecture from the ground up, and redeploys production infrastructure.
In response to the vector used in the attack, Cursor took aggressive structural steps to sever ties with the compromised environment. The company permanently transitioned all AI model inference traffic away from Fireworks AI. Furthermore, Cursor published explicit remediation guidelines directed at any customers who may have installed or updated the affected extension during the 11-hour exposure window, advising them to inspect their local SSH configurations, remove unauthorized keys, and rotate any developer credentials that might have been locally accessible.
Broader Industry Context: The "Black October" Wave of AI Infrastructure Attacks

The security incident at Cursor does not stand as an isolated event. Cybersecurity firm Wiz has characterized the broader cluster of attacks targeting AI development infrastructure throughout October as "Black October." This coordinated wave of campaigns specifically targeted the foundational layers of modern AI development platforms to compromise authentication tokens, harvest proprietary models, and hijack computing resources.
The primary motivation behind many of these incidents has been the unauthorized execution of AI inference workloads, which generates massive operational costs for victims, alongside aggressive attempts at AI model distillation—where attackers query models repeatedly to replicate their capabilities.
Numerous prominent AI-adjacent platforms fell victim to similar tactics during the same timeframe. Railway suffered a significant breach where attackers stole customer authentication secrets, subsequently leveraging them to rack up approximately $740,000 in fraudulent AI computing charges. Similarly, OpenCode disclosed compromised API keys and rampant unauthorized AI usage. Code-assistance platforms like Warp also reported security disclosures involving the exposure of AI agent logs and user telemetry data.
Fireworks AI itself was subjected to multiple scrutinized security events in October. Alongside the key exposure that impacted Cursor, Fireworks disclosed a separate incident in which attackers gained access to internal development environments via a compromised third-party cloud provider. The ripple effects of the Fireworks compromises extended to other high-profile enterprise customers, such as Notion, which acknowledged that AI chat transcripts belonging to a small number of workspaces had been exposed as a result of the vendor’s security lapse.
Implications for the AI Development Ecosystem
The Cursor breach and the broader "Black October" campaign serve as a stark wake-up call for the software development industry regarding the security posture of AI tooling. As developers increasingly integrate autonomous coding agents, cloud-based inference engines, and third-party extensions into their daily workflows, the attack surface expands exponentially.
Historically, software supply chain security has focused heavily on traditional package managers like npm, PyPI, or Maven. However, the Cursor incident demonstrates that specialized distribution channels—such as IDE extension marketplaces, AI model training buckets, and shared telemetry logs—represent high-value targets for sophisticated threat actors.
The inclusion of unredacted API credentials within historical coding-agent traces highlights a persistent challenge in AI data engineering: the accidental logging of sensitive data during model training data collection. Even when platforms offer privacy controls like Cursor’s Privacy Mode, human error or automated pipeline oversights can leave sensitive tokens exposed in intermediate storage tiers.
As the industry moves forward, cybersecurity analysts expect stricter regulatory scrutiny, enhanced data sanitization and automated redaction protocols for AI training datasets, and more rigorous code-signing and verification requirements for IDE extensions. For developers, the incident reinforces the critical necessity of adhering to the principle of least privilege, isolating local development environments, and maintaining continuous monitoring of API key usage across both first-party and third-party services.






