Fireworks AI widens a breach to model weights and datasets
A second disclosure on 9 November says attackers reached billing data, secrets and incomplete model weights for some customers, after reaching Fireworks through a compromised cloud provider.
Fireworks AI, an LLM inference platform used by Notion, Cursor and Harvey among more than 10,000 customers, widened its breach disclosure on 9 November. Attackers who reached its development infrastructure through a compromised third-party cloud provider on 23 October got into resource metadata, billing and payment history, and, for a limited set of accounts, stored secrets, datasets and incomplete model weights.
Fireworks says payment card numbers were not affected because it does not store them, and it has found no evidence attackers reached production inference traffic or training data outside the development environment. It has revoked and replaced credentials, restricted network pathways to the affected systems, and temporarily disabled self-serve dataset uploads and downloads while it works with Palo Alto Networks’ Unit 42 on the investigation. Fireworks attributes the activity to a financially motivated group running “LLMjacking”, reselling access to stolen inference credentials through illicit routing gateways.
Why it matters: if you store API keys, secrets or training data with Fireworks, rotate everything there now and check your usage and billing logs for activity you didn’t generate, the same advice Railway, Northflank and Trigger.dev customers got after similar breaches this autumn. An inference provider holding your model weights and datasets is a new kind of blast radius beyond the usual leaked-API-key story.
The caveat: Fireworks has not said how many accounts had secrets, datasets or weights exposed, only that affected customers were notified directly.