Fireworks AI gets burned twice as customer data exposure widens
First reported by The Register ·
Your AI model training data and billing information are now at higher risk of exposure. You will need to review access controls and data handling practices for all AI services you use. This incident highlights the need for enhanced due diligence on third-party cloud providers in the AI supply chain.
Fireworks AI has disclosed a second cyberattack in October, resulting in a broader exposure of customer data. The initial breach on October 6 involved unauthorized access to internal credentials and customer API keys. The subsequent attack on October 23 exploited a vulnerability in a third-party cloud provider used for Fireworks' research and development, allowing attackers to access Fireworks' development systems. This second incident led to the exposure of billing records, resource metadata, and a subset of incomplete model weights and datasets for certain customers. While the company stated that payment card numbers were not affected, user conversations with Notion AI, which utilized Fireworks, may have contained identifying details. Fireworks is working with affected customers like Cursor and Harvey, and has engaged Palo Alto Networks' Unit 42 for remediation.
This incident signals a concerning trend in AI security, where the interconnectedness of services and reliance on third-party infrastructure amplify the impact of breaches. LLMJacking, a technique combining unauthorized access to AI models with financial motives, appears to be an emerging threat vector. The exposure of incomplete model weights and datasets, while not directly stolen, suggests that attackers are probing for valuable IP and potentially seeking to disrupt ongoing training or inference operations.
The widening scope of the data exposure, including billing information and potential conversational data, raises significant compliance and privacy concerns for Fireworks' customers. Companies utilizing Fireworks' platform, such as Notion, Cursor, and Harvey, now face the reputational and operational challenges of managing customer trust following a data incident. The reliance on a single cloud provider for R&D infrastructure also highlights a potential single point of failure that could be targeted by sophisticated adversaries.
AI-written summary. May contain errors.