Model hosting platform Fireworks has confirmed that it has suffered a new round of hacker attacks. The leaked information included billing information, model weights and data sets of some customers. The company is investigating along with Unit 42, a unit of Palo Alto Networks.
Pay attention to the words "new round". This is not the first time something has happened to it.
For an inference platform, the leakage of billing information is troublesome, and the leakage of weights and data sets is fatal. Customers put fine-tuned models here to run, and the weights are the core assets they spend money and time to train. The data set often also contains business data.
Legal AI company Harvey, a high-profile Fireworks customer, said its client data was not affected. It's unclear whether other customer and end-user data was compromised.
The timing was terrible. Fireworks' annualized revenue has reached US$1 billion in July this year. It originally planned to launch a new round of financing this week with a target valuation of at least US$30 billion. Some big clients who pay millions of dollars a year are considering moving, people familiar with the matter said.
Migration is not difficult between inference platforms, which is a characteristic of this industry. What everyone sells is open source model hosting and acceleration, and the interfaces are basically compatible. You can just change the address and key by another company. Usually customers rely on price, latency and stability to stay, but now there is an additional one: can you keep an eye on my weight?
This is a reminder if your team hosts fine-tuned models on any third-party platform. Ask the supplier three questions: whether the weight is encrypted at the storage layer, who has the permission to read it, and how long it will take to notify the customer after an incident occurs. If you can’t answer, you should have a backup plan.