AI inference giant Fireworks, which generates $1 billion in annual revenue, has been hacked twice in succession, leading major clients to consider cutting ties with the firm.

Beating AI News Flash: AI inference and model training platform Fireworks has suffered two consecutive hacker breaches. On November 9, the company admitted that some customers’ billing information, datasets, and unfinished model weights had been accessed without authorization. Two clients interviewed by Forbes said they have moved most of their business away from Fireworks. The first breach occurred in early October, when hackers stole Fireworks’ internal credentials to obtain API keys used by some customers for model evaluation. At the time, the company stated it had not detected any unauthorized access to customers’ models or datasets. On October 24, hackers gained access to Fireworks’ research and development machines via a compromised third-party cloud service provider, running unauthorized programs that further impacted customer data. One client interviewed by Forbes complained that Fireworks has faced increasingly frequent hacker attacks and poor communication with customers. Some major clients paying millions of dollars annually to Fireworks are considering switching vendors. Fireworks reported $1 billion in annual revenue in July this year, with a valuation of $17.5 billion. The company originally planned to launch a new financing round soon, targeting a valuation potentially exceeding $30 billion. The client exodus could put pressure on this fundraising. Fireworks said the attackers are likely linked to a cybercriminal group that steals and resells large language model (LLM) inference credits. No impact has been found on core services such as online inference, and no payment card numbers have been leaked. The company has suspended some data upload and download functions, and enlisted Unit 42, Palo Alto Networks’ cybersecurity division, to assist with the investigation.