Fireworks AI has disclosed further developments in its investigation of a security breach initially reported last month. The breach, which was originally thought to be limited, has now been found to affect customer billing information, secrets, datasets, and model weights for a select number of accounts.
Unauthorized Access and Initial Findings
In an announcement dated October 10, 2026, Fireworks AI reported detecting unauthorized use of its internal credentials. This breach allowed access to environment variables used in evaluation jobs for certain customer accounts. The unauthorized activity was first identified on October 6, 2026, at 9 a.m. PDT. Initially, the company asserted that only environment variables, such as API keys and tokens, were compromised, with no evidence of access to customer models, datasets, or personal data.
In response, Fireworks AI revoked compromised credentials, rotated service account tokens, and updated affected API keys. Despite these measures, further investigation revealed a more extensive breach.
Expanded Breach and Impact
The latest update indicates that attackers gained access to Fireworks’ development environment via a third-party cloud provider on October 23, 2026. This entry point was utilized to run unauthorized workloads, further impacting customer data. As of October 25, 2026, no additional unauthorized access has been detected.
The compromised data includes resource metadata, billing and shipping addresses, invoice history, account balances, and incomplete model weights. Importantly, payment card details were not affected as they are not stored by Fireworks AI. Customers impacted by this breach have been notified.
Security Measures and Industry Impact
In response to the breach, Fireworks AI has implemented enhanced security measures, including credential revocation, system isolation, and improved access controls and monitoring. The breach has been linked to a financially motivated cybercriminal group engaged in a form of cybercrime known as LLMjacking. This involves hijacking large language models through stolen credentials and selling access to other threat actors.
Google has noted a rise in LLMjacking incidents targeting AI infrastructure across various sectors in North America, South America, and the Asia Pacific.
Notion’s Involvement and Customer Advisory
Notion, a popular productivity application, was also impacted by the Fireworks AI breach. The company disclosed that less than 1% of its workspaces were affected, with unauthorized access potentially extending to customer data. Initial assessments suggested the breach impacted Notion AI’s synthetic testing data, but further analysis revealed involvement of chat transcripts used in the Insights feature.
Notion has since disabled the Insights feature and ceased using open-weight models from Fireworks AI. Affected customers are advised to review their data and update any sensitive information that may have been compromised during the breach.
