• Sources: primary, discussion, discussion
  • Thread: Fireworks AI cloud provider breach and customer data exposure
  • Summary: Notion says that Fireworks AI, which it used for inference and for classification in its Insights feature, confirmed on 2026-11-08 that a stolen dataset included Notion AI chat transcripts with no user or workspace labels, affecting under 1% of workspaces. Notion disconnected from Fireworks on 2026-11-03 and advises rotating any credentials pasted into Notion AI. BleepingComputer reports that the attackers used a cloud provider breach on 2026-10-23 to reach Fireworks R&D machines and took billing data, resource metadata, and partial model weights and datasets. The incident follows an earlier Fireworks breach of API key secrets tied to LLMjacking.
  • Why it matters: If you send prompts or fine-tuning data to a hosted inference provider, its subprocessors are part of your attack surface. Rotate any secrets that users pasted into chat.
  • Background: Inference platforms such as Fireworks host open-weight models and customer fine-tunes behind an API, so they hold prompts, datasets and weights from many tenants. LLMjacking means stealing API keys or cloud access to resell model usage. Several AI infrastructure and agent platforms disclosed breaches aimed at stored AI keys in October and November 2026.