Fireworks AI Breach Widens to Model Weights and Notion Chats
By @sharedot · · 8 pages
- Cybersecurity
- Data Breach
- Llmjacking
- Fireworks AI
- Notion
Fireworks AI says a second October attack exposed customer datasets, incomplete model weights, billing records, and Notion AI chat transcripts.
A Second Breach the Company Had Ruled Out
Fireworks AI disclosed Monday that its investigation into an incident first reported in early October uncovered further malicious activity, affecting account and billing information alongside secrets, datasets, and incomplete models for a 'limited set of accounts.'
The company had initially detected unauthorized use of internal credentials on October 6, 2026, and said at the time that there were no signs customer models, datasets, training data, or personal information were accessed. The latest bulletin substantially broadens that assessment.
The Register reports the update pins the wider exposure on a compromised third-party cloud provider hosting Fireworks research and development machines, which attackers used to reach development systems and run an unauthorized workload.
Attackers Arrived Through a Third-Party Cloud
The second intrusion began on October 23, 2026, when an unauthorized third party compromised an unnamed third-party cloud provider that hosted Fireworks' R&D machines.
According to the company's bulletin, on October 24 the attacker used that foothold to reach Fireworks development machines and run an unauthorized workload that led to additional customer data impact.
Fireworks said that as of October 25 it had not detected further evidence of successful unauthorized access on the platform customers use.
This supply-chain path is the surprising reversal: an AI infrastructure provider was breached not directly but through the cloud hosting its own research machines, the same pattern that made the original credential theft seem contained.
What the Attackers Could Reach
Fireworks grouped the exposed data into three categories: resource metadata such as identifiers, display names, and creation times; billing information including addresses, invoice and payment history, account balances, and credit usage; and, for certain customers, a subset of incomplete model weights and datasets.
The company said payment card numbers are not stored and were not impacted. CEO Lin Qiao pushed back on the idea that model theft was the goal, telling The Register that 'from what we see attackers are not after weights' and that most downloads were partial and stopped.
Bloomberg reported the company is also working with customers including Cursor and Harvey to assess the impact.
Notion's AI Chats Are Caught in the Blast Radius
Notion confirmed that the Fireworks incident affected fewer than 1% of its workspaces.
Its Insights feature, which helps teams understand AI usage, processed the affected conversations through Fireworks, and reviews ultimately identified customer chat transcripts — reversing Notion's initial finding that only synthetic testing data was exposed.
For the vast majority of impacted workspaces the transcripts were 'unattributable,' meaning they lacked labels tying chats to authors or workspaces and contained no tool call results.
The company cautioned, however, that anything typed into a chat — passwords, API keys, identifying details — could still expose people or businesses. Notion has disabled Insights and stopped serving open-weight models through Fireworks.
The Motive: LLMjacking, Not Weight Theft
Fireworks said evidence indicates links to a financially motivated cybercriminal group aimed at selling unauthorized access to large language models through illicit routing gateways — an emerging threat known as LLMjacking, where stolen credentials hijack victims' LLM capacity while the legitimate owner pays the bill.
Google warned last week that such attacks are increasingly targeting AI infrastructure in North America, South America, and Asia Pacific across financial services, retail, technology, and transportation, and The Hacker News notes Google has said LLMjacking is costing some enterprises up to $1 million in cloud charges.
The Register reports Fireworks worked with Palo Alto Networks' Unit 42 to remediate.
Containment and What Affected Customers Should Do
Fireworks revoked and rotated compromised internal credentials and service account tokens, isolated affected systems to prevent lateral movement, tightened access controls, and improved logging and monitoring.
Notion has contacted affected customers with instructions to review impacted data securely inside their workspaces; customers who received no notification have not been identified as affected.
Anyone who pasted sensitive information, passwords, API keys, or other credentials into Notion AI or affected Fireworks accounts during the attack window should rotate them immediately. Notion's investigation remains open, with further updates available through Fireworks' security bulletins.
Sources
- thehackernews.com › Fireworks AI Confirms Wider Breach Affecting Customer Datasets, Model Weights and Notion AI Chats
- theregister.com › Fireworks AI gets burned twice as customer data exposure widens
- cybersecuritynews.com › Notion Confirms AI Chat Transcripts Exposed in Fireworks AI Breach, Affecting Customer Workspaces