AI infrastructure provider Fireworks says hackers accessed some customers' datasets, incomplete model weights, and billing information over the past month

Your AI model weights and datasets are now at risk of unauthorized access and potential resale.

AI infrastructure provider Fireworks AI disclosed a security incident impacting some customers. Initially thought to only affect stored secrets, the breach expanded to include account and billing information. For a subset of customers, hackers also accessed datasets and incomplete model weights. The incident originated from a third-party cloud provider's security breach on October 23, which allowed unauthorized access to Fireworks' development machines on October 24. Fireworks detected the unauthorized activity on the same day and implemented remediation measures. The company is working with cybersecurity firm Palo Alto Networks Unit 42 and law enforcement. The attack appears linked to financially motivated cybercriminal groups engaging in "LLMJacking," a practice of selling unauthorized LLM access tokens. Fireworks has implemented enhanced security measures, including credential lockdown, network isolation, and stricter data access controls, and has temporarily disabled self-serve dataset uploads and downloads.

The incident highlights the increasing sophistication of attacks targeting AI infrastructure and the supply chain risks inherent in the generative AI ecosystem. LLMJacking, as described by Google Threat Intelligence, signifies a new frontier for cybercriminals seeking to monetize AI assets, potentially impacting the economics and security of AI model deployment. This breach will likely intensify scrutiny on cloud security and data protection practices within AI development platforms.

Customers affected by this breach may face significant repercussions, including intellectual property theft and competitive disadvantage if their model weights or proprietary datasets are compromised. The temporary disabling of dataset upload/download features could also disrupt ongoing development and training workflows, forcing companies to re-evaluate their operational security and reliance on third-party AI infrastructure providers.