Fireworks, an AI infrastructure startup offering high-performance inference capabilities for deploying open-source language models like Mistral, Meta's Llama, and Google's Gemma, has disclosed a second security breach occurring in October. The incident followed an earlier compromise in which threat actors obtained API keys and tokens from a limited set of customer accounts.
The company reported the initial breach to law enforcement and customers in early October after discovering unauthorized access to environment variables containing API credentials. Fireworks attributed the activity to a financially motivated cybercriminal group targeting tokens for resale through illicit routing gateways in what the company described as "LLMJacking" attacks.
Following credential revocation and remediation efforts, Fireworks identified a second incident on October 23. In this attack, threat actors accessed resource metadata including Fireworks identifiers, display names, descriptions, and creation timestamps. The breach also exposed billing details such as addresses, invoice and payment history, account balances, and credit usage information for some customers.

Fireworks clarified that payment card numbers remained secure because the company does not retain such information. The attackers also obtained incomplete model weights and datasets for certain customers.
"A third-party cloud provider that hosted Fireworks R&D machines experienced a security incident on October 23. On October 24, the unauthorized third party used it to reach Fireworks development machines and then run an unauthorized workload that led to additional customer data impact," the company stated in a Monday advisory.
Fireworks detected the unauthorized activity on October 24 and executed incident response procedures to address the breach. The company confirmed that as of October 25, no further successful unauthorized access attempts targeting the customer-facing Fireworks platform had been detected.
Remediation and Investigation
In response to the breach, Fireworks has implemented several security measures. The company revoked and replaced all credentials and service accounts, eliminated unauthorized network pathways, quarantined compromised systems, and restricted access to data storage and analytics infrastructure to prevent unauthorized retrieval of stored information.
Fireworks is collaborating with Palo Alto Networks' Unit 42 incident response team to strengthen defenses against future attacks and conduct a comprehensive investigation into the incident's scope and consequences.
Broader Context on LLMJacking Costs
The timing of Fireworks' disclosure comes one week after researchers from Google's Threat Intelligence Group and Mandiant released findings on large language model hijacking attacks. Their research documented that LLMjacking campaigns are generating unprecedented costs for affected organizations, with fraudulent cloud AI charges reaching as high as $1 million when attackers exploit stolen API keys to make unauthorized calls.