Stolen Fireworks AI Key Enables Trojanized Claude Code Extension

Introduction
A supply chain compromise affecting Anthropic's official Claude Code extension on the Cursor extension marketplace has been traced to the theft of an access key originally issued to Fireworks AI. Cursor disclosed on Friday that an attacker obtained a key exposed during a breach at Fireworks, then used it to access read-only data and ultimately deliver a trojanized version of the popular AI coding assistant extension. The incident underscores how a single leaked credential can cascade across interconnected AI development platforms, putting downstream developers at risk.
Attack Chain: From Fireworks Breach to Cursor Marketplace
According to Cursor's update, the attacker first acquired an access key that Cursor had provisioned for Fireworks AI. That key was subsequently exposed—likely through the Fireworks breach—and the attacker leveraged it to copy data from a read-only scope. While the exact mechanics of how read-only access translated into publishing a modified extension remain under investigation, the outcome is clear: a malicious version of the Claude Code extension was made available to developers installing from Cursor's marketplace.
This pattern mirrors classic supply chain attacks: compromise a trusted partner, pivot using legitimate credentials, and poison a widely consumed artifact. Because the extension bore Anthropic's branding and was hosted on Cursor's official marketplace, developers had little reason to suspect foul play.
Why This Matters for AI-Native Development Workflows
AI coding assistants like Claude Code are rapidly becoming integral to daily engineering workflows. They often run with elevated permissions—accessing source code, terminals, and cloud credentials. A trojanized extension could exfiltrate intellectual property, inject backdoors into generated code, or harvest authentication tokens for cloud environments.
The blast radius extends beyond individual developers. CI/CD pipelines that automatically install or update extensions could propagate the compromise to build systems, staging environments, and even production artifacts. Organizations that rely on Cursor or similar AI-enhanced IDEs must treat extension integrity as a critical control point.
Immediate Mitigation Steps for Defenders
- Audit installed extensions: Verify the hash and publisher signature of every Claude Code extension instance across developer machines and CI runners. Remove any versions not matching the official Anthropic release.
- Rotate potentially exposed credentials: If your organization uses Fireworks AI or shares keys with partners, rotate all API keys, service accounts, and personal access tokens immediately.
- Enforce extension allow-listing: Configure IDE and marketplace policies to permit only vetted, hash-pinned extensions. Block automatic updates until manual verification.
- Monitor for anomalous extension behavior: Deploy endpoint detection rules that flag unexpected network connections, file system access, or process execution originating from extension hosts.
- Demand transparency from vendors: Request detailed incident timelines, scope of data accessed, and remediation proof from both Cursor and Fireworks AI.
Grivyonx Expert Analysis
This incident reveals a systemic weakness in the emerging AI tooling supply chain: trust is transitive, but credential hygiene is not. Cursor issued a key to Fireworks; Fireworks failed to protect it; attackers weaponized it to poison an Anthropic artifact hosted on Cursor. Each party likely met baseline security requirements, yet the composite system failed. Defenders must move beyond vendor questionnaires and adopt continuous, cryptographic verification of every artifact—extensions, container images, model weights—that enters their build graph. Automated policy enforcement, not periodic audits, is the only scalable answer.
Conclusion
The trojanized Claude Code extension serves as a stark reminder that the AI development stack is only as strong as its weakest credential. As AI assistants gain deeper access to codebases and infrastructure, the incentive for supply chain attacks will only grow. Organizations should treat extension and plugin ecosystems with the same rigor applied to third-party libraries and container images. Grivyonx helps teams embed automated supply chain verification into DevSecOps pipelines so that trust is continuously validated, not assumed.
- #supply chain attack
- #extension security
- #Fireworks AI
- #Cursor
- #Claude Code
- #credential theft
- #DevSecOps



