Skip to content
SM Secure Blog

How a Stolen Fireworks AI Key Triggered a Cursor Breach and Trojanized Claude Code Extension

Sagar Majumdar

Sagar Majumdar

Modern AI coding assistants promise effortless velocity, wiring frontier language models directly into local integrated development environments. Yet beneath that smooth autocomplete experience lies a sprawling, interconnected supply chain of inference providers, extension marketplaces, and telemetry storage buckets. This weekend, security disclosures from Cursor revealed how a single compromised partner credential cascaded into a multi-stage supply chain attack, exposing customer data and delivering trojanized extensions directly onto developer workstations. The intrusion originated upstream at third-party AI inference provider Fireworks AI. According to Cursor's incident update, threat actors obtained an access key that Cursor had originally issued to Fireworks AI for operational integration. Armed with that stolen credential, the attackers pivoted into a read-only storage bucket associated with Cursor, exfiltrating historical coding-agent traces from sessions where Privacy Mode was disabled, archived non-production source code repositories, customer support conversations, and unredacted API credentials that users had previously shared during support interactions. Rather than stopping at passive data theft, the attackers leveraged the harvested secrets to escalate deeper into Cursor's internal ecosystem. By compromising extension publishing pipelines, they injected malicious builds of Anthropic's official Claude Code extension, specifically versions 2.1.312 and 2.1.313, into Cursor's marketplace cache. For roughly eleven hours on October 29, developers installing or updating the trusted extension unwittingly downloaded a trojanized package engineered to plant persistent SSH backdoors on their machines. Cursor estimates that up to 5,000 developer endpoints were exposed before the tainted builds were purged and infrastructure was rebuilt. What makes this breach particularly instructive for engineering and security teams is how cleanly it illustrates the hidden attack surface of AI developer tooling. Developer workstations sit at the very center of corporate trust boundaries, routinely holding production cloud keys, private repository tokens, and signing certificates. When an IDE marketplace or an external inference vendor suffers a credential leak, the blast radius extends immediately onto the local machines of engineers who rely on those tools daily. Furthermore, the presence of unredacted user API keys inside stored agent traces and support logs highlights a persistent hygiene failure across fast-moving AI platforms. In response, Cursor rotated its exposed credentials, migrated inference traffic away from Fireworks AI, and rebuilt its extension marketplace infrastructure while confirming that sessions operating under Privacy Mode remained unaffected. However, platform-level remediation alone cannot insulate organizations from upstream supply chain compromises. Engineering teams should enforce strict secret-scanning across local environments and support tickets, treat third-party IDE extensions with the same vetting rigor applied to production software dependencies, and rotate any credentials that may have passed through unencrypted coding-agent telemetry. As AI tooling becomes deeply embedded in daily software delivery, securing the inference and extension supply chain is no longer optional; it is foundational endpoint defense.

Comments

No comments yet. Be the first to add one.