A breach at a third-party cloud provider gave attackers a path into Fireworks AI’s development machines, where they ran an unauthorized workload that led to additional customer-data exposure. One affected dataset included Notion AI chat transcripts used for the Insights feature. Notion said fewer than 1% of its workspaces were affected and clarified that the incident was at its vendor, not in Notion’s own systems.
4
8
How the intrusion unfolded
Fireworks said the cloud provider hosting its research-and-development machines experienced a security incident on October 23. On October 24, an unauthorized party used that path to reach Fireworks development machines and run a workload. Fireworks detected the activity that day.
4
14
The available disclosures describe the route through the cloud provider and development environment, but do not establish the precise technical exploit or every step by which the attackers reached particular data. The key point is that access to a vendor’s development environment led to additional customer-data impact.
14
What data was exposed
Fireworks’ expanded disclosure included resource metadata and billing information, including billing and payment history. It also said datasets and incomplete model weights were affected for some customers. Fireworks said it does not store payment card numbers.
4
9
34
For Notion, the affected material was chat transcripts used in classification jobs for Insights, a feature designed to help customers understand their AI usage. Notion reported that fewer than 1% of workspaces were affected; for the vast majority of those, the transcripts were not attributable to a specific user or workspace.
3
4
Notion’s response timeline
- November 2: Fireworks told Notion that a small dataset had been accessed. Notion’s initial review found synthetic testing data, not customer chats.
8
- November 3: Notion disconnected from Fireworks and deleted its data held in the vendor’s systems.
5
- November 6: Fireworks warned Notion that customer data might also have been accessed.
5
- November 8: Fireworks confirmed that the dataset included Notion AI chat transcripts used for Insights.
8
- November 9: Notion began contacting affected customers.
15
- November 11: Notion revoked exposed API keys found in affected user messages.
8
Notion also disabled Insights and stopped serving open-weight models through Fireworks, according to reports on the company’s response.
10
11 Affected customers were advised to treat passwords, API keys, and other credentials entered into the assistant as potentially exposed and rotate them.
21
Investigation and attribution
Fireworks engaged Palo Alto Networks’ Unit 42, notified law enforcement, and restricted some affected functionality.
34 The company said it believed the attackers sought to resell unauthorized model access through illicit routing gateways, a practice known as “LLMJacking.” That is Fireworks’ assessment of the motive, not conclusive identification of an attacker.
7
The cloud-provider incident was disclosed as a second Fireworks breach in October, following an earlier incident identified on October 6.
7 The two incidents should not be conflated: the later route described here began with a third-party provider and reached development machines on October 24.
4
7
What the incident shows about AI supply-chain risk
A company’s security exposure can extend beyond its own systems. When an AI service sends data to an outside provider for processing, the provider’s infrastructure and development environment become part of the broader risk picture. Notion’s disclosure is an example: the reported entry point was at Fireworks’ cloud provider, while the affected conversations belonged to a Notion feature.
4
8
For teams evaluating AI services, the incident is a reminder to ask what data a vendor or subprocessor retains, where it is processed, and how quickly access can be cut off. Avoid putting credentials into prompts when possible; if a secret may have appeared in an exposed transcript, revoke or rotate it. The reported exposure also reinforces the value of limiting access to datasets, credentials, and development systems to what is needed for a task.