Fireworks AI disclosed a two-part breach on November 10, 2026. Between November 7 and 8, an attacker compromised an internal developer tooling environment and exfiltrated two customer AI API keys. Within hours, Palo Alto Unit 42 confirmed active LLMjacking using the stolen credentials: the attacker ran approximately 4.2 million tokens per hour on Fireworks-hosted models before detection. The breach adds to a growing pattern of AI API key theft and rapid monetization that has emerged over the past 60 days.
What happened
Fireworks AI's status page advisory describes a compromise of internal developer tooling between November 7 and 8. The attacker accessed an environment that held customer API credentials and exfiltrated keys belonging to two Fireworks customers. Fireworks detected the unauthorized access and engaged Palo Alto Unit 42 for incident response. Affected customer accounts were suspended pending key rotation. The root cause of the initial tooling compromise is under active investigation.
LLMjacking confirmed: 4.2 million tokens per hour
Unit 42's incident response team confirmed that the stolen API keys were used for LLMjacking: running high-volume, unauthorized inference workloads on the victim accounts. The attacker sustained approximately 4.2 million tokens per hour before Fireworks detected and suspended the accounts. At standard Fireworks pricing for a mid-tier model, that rate translates to several thousand dollars per hour in compute costs billed to the victim customers.
The LLMjacking arc is accelerating
The Fireworks AI incident is the third confirmed LLMjacking case in 60 days. Prior incidents involved an unnamed cloud AI provider and an AI startup in October 2026. The pattern is consistent: steal API keys from a development or CI/CD environment, start high-volume inference jobs within hours, generate revenue through resale of model access or consume compute for downstream tasks before the victim notices. The monetization window in each confirmed case has been under 24 hours.
Why AI API keys are high-value targets
Unlike cloud provider credentials, AI API keys often receive less stringent security treatment. They are frequently stored in developer dotfiles, hardcoded in scripts, committed to repositories, or retained in CI/CD secrets managers without rotation. The access they grant is immediately liquid: an attacker can convert a stolen AI API key into compute value within minutes, with no need for the data exfiltration or lateral movement that cloud credential theft typically requires to generate revenue.
Action items
- Rotate all AI API keys across providers now, particularly any keys that have been in use for more than 30 days or that live in developer environments, dotfiles, CI/CD pipelines, or shared secrets managers.
- Enable spend alerts and usage anomaly detection on all AI API accounts. Set hard rate limits and spending caps. A 4.2 million token per hour anomaly should trigger an alert within minutes, not hours.
- If you are a Fireworks AI customer, verify with Fireworks whether your keys were among those exfiltrated. Even if you were not affected by this incident, treat AI API keys with the same rotation and monitoring discipline you apply to cloud provider credentials.
Gigia Tsiklauri is a Security Architect and founder of Infosec.ge. Get in touch if you need help building AI API key governance or responding to an AI credential incident.