What happened
On November 9, 2026, Fireworks AI expanded its breach disclosure: attackers accessed account/billing info and, for a limited group, secrets, datasets and incomplete model weights. Evidence links the activity to criminals reselling unauthorized LLM access via LLMJacking gateways. Palo Alto Unit 42 engaged. Inference/Training SDK/Nexus use zero data retention.
Why it matters
This is a real-world LLMJacking campaign against a major AI inference provider that progressed from credentials to model artifacts and datasets. It shows the monetization model (stolen cloud identity tokens resold via illicit gateways) and that development machines / third-party hosting are the weak link.
Attack vector
Financially motivated LLMJacking group. After an initial October 6 credential-secret exposure (environment variables used in evaluation jobs), a second intrusion on October 24 used an October 23 security incident at a third-party cloud provider hosting Fireworks R&D machines as a route into development machines.
Affected systems
Fireworks AI platform (inference, Batch Inference, dataset-based training, Training SDK, Nexus)
Mitigation
Rotate/replace revoked Fireworks API keys; rotate exposed third-party credentials with their issuing providers; review logs for unrecognized keys/users/evaluators/training jobs/datasets; self-serve dataset uploads/downloads are temporarily disabled.