AI hosting firm Fireworks AI says a cloud provider breach exposed customer data and some model weights
Fireworks AI widened its breach disclosure on November 9, saying attackers who compromised a third-party cloud provider reached billing data for all affected accounts and, for a smaller group, secrets, datasets and incomplete model weights.
Fireworks AI, an AI model-hosting platform backed by Nvidia, told customers on November 9 that a security incident it had previously disclosed was broader than first described. The company said attackers had compromised a third-party cloud provider that hosted some of its research and development machines, then used that foothold to reach Fireworks systems and run an unauthorized workload, according to reporting by The Register and The Hacker News.
According to Fireworks, the intrusion traced back to late October, when the third-party provider was first compromised before attackers moved into Fireworks' own development environment. The company said this was separate from an earlier incident in early October, in which someone used internal credentials to read environment variables tied to some customer accounts.
In the November 9 update, Fireworks said attackers reached account and billing information for all customers affected by the incident. For a smaller subset of accounts, it said the exposure also included secrets, datasets, and what it described as 'incomplete' model weights, without detailing what portion of any model was reachable. The company said payment card numbers are not stored on its systems and were not affected.
Notion, which routes some AI features through Fireworks, said separately that AI chat transcripts used by its Insights classification tool were part of the exposed data, affecting fewer than 1% of its workspaces. Notion said it has disabled Insights and open-weight model routing through Fireworks, stressed that its own systems were not breached, and advised affected teams to rotate any secrets that may have appeared in those chats.
Fireworks CEO Lin Qiao said the company's working theory is that the attackers were financially motivated and likely tied to so-called LLMjacking, where stolen credentials are used to run AI inference at someone else's expense rather than to steal proprietary model weights. Fireworks said it has notified law enforcement and brought in Palo Alto Networks' Unit 42 for forensic analysis.
As a precaution, Fireworks has disabled self-serve dataset uploads and downloads for batch inference and dataset-based training while the investigation continues, and the company says it has stopped serving some open-weight models through its platform. It is urging customers to rotate any secrets or API keys stored on the service and to review billing activity for unexpected usage.
The incident is the latest in a string of breaches touching AI infrastructure providers in 2026, and it illustrates a growing risk for companies that plug third-party AI hosting and inference services into their own products: a compromise upstream can expose customer data several layers removed from the original breach. Fireworks says its investigation is ongoing and the full scope of what was accessed has not yet been finalized.
For businesses relying on outside AI vendors, the episode is a reminder to track which secrets and datasets get handed to third-party platforms, and to have a plan for rotating credentials quickly when a vendor discloses an incident.
From Spaqitech & Spaqi Consulting
Sources
- The Register: Fireworks AI gets burned twice as customer data exposure widens
- The Hacker News: Fireworks AI Confirms Wider Breach Affecting Customer Datasets, Model Weights and Notion AI Chats
This summary was written with AI assistance from the sources above. The image is illustrative and may not show the events described.