Fireworks AI recently revealed a significant security breach, compromising various customer account details, billing information, datasets, and incomplete model weights. In an update on November 9, 2026, the company outlined how attackers infiltrated its development systems through a third-party cloud provider, broadening the scope of the incident initially reported in October.
Initial Breach Details
Originally, Fireworks AI’s first disclosure focused on unauthorized access to secrets used in evaluation tasks across a limited number of accounts. Subsequent investigations have revealed that the breach included access to a wider array of customer data, such as unfinished models and datasets from specific clients.
The breach started on October 23, when a security incident occurred at the cloud provider managing Fireworks’ research and development infrastructure. The following day, an unauthorized entity exploited this breach to access Fireworks’ development machines, executing unauthorized workloads and causing further data exposure.
Response and Mitigation Efforts
Upon detecting this activity, Fireworks AI immediately responded by sealing critical vulnerabilities and has since reported no further unauthorized access on its customer platform as of October 25. The company has not disclosed the identity of the cloud provider or the specific vulnerabilities involved.
Fireworks has involved law enforcement and Palo Alto Networks Unit 42 for forensic analysis. The company has taken actions such as replacing credentials, reducing permissions, isolating affected systems, and enhancing storage access controls to mitigate future risks.
Impact on Customers and Future Precautions
The breach affected internal applications, exposing resource identifiers, billing addresses, and credit usage information. However, payment card numbers remained secure as Fireworks does not store these details. Some customers experienced unauthorized access to subsets of their incomplete model weights and datasets, though no complete production models were reported stolen.
Fireworks believes the attackers are financially motivated, targeting large language models to sell unauthorized access via illicit networks, a method known as LLMJacking. The company has disabled self-service dataset uploads and downloads, while ensuring training environments remain secure.
Customers have been advised to replace revoked keys and rotate exposed third-party credentials. They should also scrutinize access logs and unfamiliar activities to prevent further unauthorized access.
Fireworks AI’s ongoing investigation aims to enhance its cybersecurity posture and prevent similar incidents in the future, emphasizing the importance of robust cloud security practices.
