Threat feed

Fireworks AI breach cluster

Fireworks AI breach spread to Cursor, Uber and Notion through vendor keys and credential-bearing coding-agent traces

Posted · Spectrum Delta Research Team

Summary

An unattributed attacker used an access key that Cursor issued to its AI inference vendor Fireworks AI, exposed in that vendor's October breaches, to copy coding-agent traces from a shared storage bucket and harvest unredacted credentials from them, reaching Cursor's internal SaaS tools and planting trojanized Claude Code extension builds for about 11 hours, Cursor disclosed (A/1). Fireworks' widened disclosure of November 9 adds account and billing data, stored secrets, datasets and incomplete model weights, and the same breach cluster reached Uber, which is notifying about 150,000 drivers, and Notion, which lost AI chat transcripts from under one percent of workspaces. Fewer than 0.01 percent of Cursor customers had API credentials exposed in the traces, which came from coding sessions not using Privacy Mode. Railway, OpenCode, Warp and Fireworks disclosed similar AI-key thefts between October 6 and 23, which Railway assesses to be a series of attacks on platforms that hold AI keys.

Breakdown

Countries

  • United States

Targets

OrganizationsIndustries

  • Organization: Cursor
  • Organization: Fireworks AI
  • Organization: Uber
  • Organization: Notion
  • Industry: AI developer platforms

Technical exposure

  • Vendor-issued access keys exposed in the vendor's own environment
  • Coding-agent traces stored unredacted in shared storage buckets
  • AI coding sessions run without Privacy Mode retaining credentials
  • Credentials shared with support teams stored in SaaS tools
  • Marketplace distribution caches writable with a stolen credential

Adversary

Unattributed (financially motivated per Fireworks and Uber disclosures) · Cursor, Fireworks AI (A/1)

Procedures (4)

  1. Vendor access key used to copy shared trace bucket

    Logs

    • Cloud storage audit logs · Bulk reads of a read-only bucket used for vendor data sharing, authenticated with the vendor's access key, outside expected transfer patterns (S3 data events or equivalent object-storage audit trail)
  2. Credentials harvested from coding-agent traces

    • T1552.001Unsecured Credentials: Credentials In Files

    Logs

    • Storage access logs · Reads of coding-agent trace storage covering sessions without Privacy Mode; traces containing API credentials that were never redacted
    • DLP scan results · Secret patterns (API keys, tokens) detected in stored AI coding-agent traces and training-data buckets
  3. Stolen credentials pivot into SaaS tools

    • T1078Valid Accounts
    • T1528Steal Application Access Token

    Logs

    • SaaS audit logs · Sign-ins with credentials found in the traces, first time for that identity or from unfamiliar sources; access to support tools, archived source repositories and customer support data
  4. Trojanized extension planted in marketplace cache

    • T1195.002Supply Chain Compromise: Compromise Software Supply Chain

    Logs

    • Extension install logs · Installs of Anthropic.claude-code versions 2.1.312 or 2.1.313 from the Cursor marketplace on October 29 (builds live about 11 hours, up to 5,000 users); extension attempts to add an SSH key to authorized_keys and create a reverse SSH tunnel

Sources

Questions, detections or sightings of Fireworks AI breach cluster? Bring them to the community.