Fireworks AI Says Cloud Provider Hack Exposed Customer Datasets and Billing Records

AI inference platform Fireworks AI has disclosed that a breach of a third-party cloud provider last month resulted in unauthorized access to customer billing records, partial model weights, and secrets — a wider impact than the company initially reported when it first disclosed the incident in October.

What Happened

Fireworks AI first detected unauthorized use of internal credentials on October 6, 2026, disclosing what it described as a limited breach affecting only secrets used in evaluation jobs. The situation escalated on October 23 when the third-party cloud provider hosting Fireworks R&D machines experienced its own security incident. Attackers used that foothold to reach Fireworks development machines the following day, running an unauthorized workload before being detected.

What Was Accessed

According to the company's November 9 update, the breach exposed more data than originally reported. Affected data includes billing and shipping addresses, invoice and payment history, account balances, and credit usage. For a limited set of accounts, attackers also accessed secrets — such as API keys stored in Fireworks environments — as well as incomplete model weights and datasets. Payment card numbers were not affected; Fireworks says it does not store them.

LLMJacking and the Threat Actor

Fireworks AI linked the attack to the same threat actor profiled in a November 2 blog post by Google Threat Intelligence Group and Mandiant, which described an LLMJacking campaign targeting AI infrastructure providers. LLMJacking attacks steal API credentials to run unauthorized AI workloads at the victim's expense — in this case, the attackers ran an unauthorized workload directly on Fireworks development machines. The company has engaged Palo Alto Networks Unit 42 for forensic analysis and notified law enforcement.

What Remains Operational

Fireworks AI says its Inference, Training SDK, and Nexus products — which operate with zero data retention by default — show no detected impact from the breach. Self-serve dataset uploads and downloads for Batch Inference and dataset-based training have been disabled while the investigation continues. The company has replaced all credentials and service accounts, isolated affected systems, and tightened access controls. Affected customers are being notified individually; those with questions can contact [email protected].

The incident joins a broader wave of AI infrastructure breaches in 2026, in which attackers have increasingly targeted AI platforms as entry points for credential theft and unauthorized compute access.