Fireworks discloses two October breaches affecting customer data
The useful fact for buyers is narrower than the AI label: secrets and billing data moved, then cloud-hosted development machines did too.
TL;DR
Fireworks told customers it found a breach on October 6 and a second incident tied to a third-party cloud provider on October 23, The Record reported. The company said affected accounts were limited and notified, but the exposed material included account and billing information, secrets, internal Fireworks data, invoices, payment history, datasets and other AI-related data. Notion said some of its customers were affected and cut off Fireworks.
Fireworks is not describing one clean intrusion with one clean boundary. According to The Record, the San Mateo AI infrastructure company said it discovered a breach on October 6, hired Palo Alto Networks Unit 42, and found that a financially motivated cybercriminal group had stolen account and billing information as well as secrets. Fireworks said the incident affected a limited number of customer accounts, and that it notified them.
The second incident is the harder one for customers to read around. Fireworks said a third-party cloud provider hosting its research and development machines was attacked on October 23. On October 24, the unauthorized party used that provider to reach Fireworks development machines and run an unauthorized workload, causing additional customer data impact. Fireworks said hackers accessed internal Fireworks data, billing information, invoices, payment history, datasets and other AI-related data.
Notion's notice shows how this becomes a customer-side problem without every customer running AI infrastructure themselves. Notion said it learned of the incident on November 2, disconnected its systems from Fireworks, and deleted all data in the Fireworks system. It later said malicious actors may have accessed potential customer data. Notion said it had used Fireworks for classification jobs in its Insights feature and as a general inference provider for open-weight models, and that it disabled Insights while investigating.
Fireworks linked the activity to the broader LLMJacking pattern covered by Google Threat Intelligence Group and Mandiant, where attackers hijack enterprise access to large language models. That's the part that should keep this out of the novelty breach bucket. The data at issue isn't only names on an invoice. It's also the material and access paths around AI workloads, where stolen credentials can become usage, cost and exposure before anyone has a neat incident boundary.
Published ·Deep Fathom