Fireworks AI breach: inference providers become a supply chain chokepoint
Fireworks AI Breach: Inference Providers as Supply-Chain Chokepoints
CSA Labs published a research note analyzing two security incidents Fireworks AI disclosed within five weeks, saying inference providers are both a collection point for customer credentials and a data processor, so a single breach can expose both stored keys and transmitted content. The first was discovered on October 6, 2026, when someone abused internal credentials to read customer API keys stored in the environment variables of some accounts' evaluation jobs; the second began on October 23 when a third-party cloud provider was compromised, and the attacker entered Fireworks development machines the next day, affecting account and billing information, keys for some accounts, datasets and incomplete model weights; Fireworks said the activity is linked to an economically motivated group engaged in LLMjacking. Notion disclosed that AI chat logs used for its Insights feature were among the leaked data, affecting fewer than 1% of workspaces, and said Fireworks had contractually promised zero data retention (ZDR) but the endpoint used actually retained data.
The original source analyzes how the Fireworks leak affected Notion users via a fourth party, and points out that ZDR commitments need to be verified feature by feature rather than remaining at the vendor level.