Fireworks AI has revealed that a security breach initially reported last month has had a more extensive impact than originally thought. The breach, which was detected on October 6, 2026, involved unauthorized access to internal Fireworks credentials, leading to the exposure of environment variables like API keys and tokens for a limited number of customer accounts. Initially, the company believed no customer models or personal information were accessed. However, further investigation showed that the breach allowed attackers to use an unnamed third-party cloud provider to infiltrate Fireworks' development environment, affecting additional customer data.
On October 24, the attackers leveraged this access to run unauthorized workloads, impacting more customer data than initially reported. Although payment card numbers were not compromised, Fireworks AI has alerted customers whose data has been affected. The company has responded by implementing security measures such as revoking compromised credentials, rotating service account tokens, and enhancing access controls and monitoring systems.
The breach is linked to a cybercriminal group engaged in a new threat dubbed 'LLMjacking', which involves hijacking Large Language Models using stolen credentials. This activity has been associated with financially motivated actors seeking to sell unauthorized access to these models. Google has also warned of increasing attacks targeting AI infrastructures across various sectors in regions including North America and Asia Pacific.
Notion, a popular productivity tool, disclosed that less than 1% of its workspaces were affected by the Fireworks breach, with the exposure of synthetic testing data and potential customer data. Notion's analysis revealed that the breach involved AI chat transcripts, although most were unattributable. The company has since disabled the affected feature and advised customers to review and secure their workspaces.
