The Fireworks AI breach exposed Notion AI chat transcripts from fewer than 1% of Notion workspaces, and Notion says that data should never have been stored at Fireworks. Notion disclosed that Fireworks, which it used as an inference provider, had given it a contractual commitment to zero data retention, and that its own investigation found data had been retained from the Fireworks endpoints it used.
The Fireworks AI Breach Began at a Compromised Cloud Provider
Fireworks says the intrusion began at a third party. In a security bulletin last updated on 9 November 2026, the company said a cloud provider hosting its research and development machines had a security incident on 23 October. On 24 October, an unauthorized party used that access to reach Fireworks development machines and run an unauthorized workload. Fireworks said it detected the activity the same day and, as of 25 October, had found no further evidence of successful access to the platform its customers use. It notified law enforcement and brought in Palo Alto Networks Unit 42 for forensic work. The internal application involved exposed resource metadata, billing details such as addresses, invoices and credit usage, and, for some customers, a subset of incomplete model weights and datasets. Fireworks said it does not store payment card numbers. The bulletin updates an earlier one from 9 October, which at the time listed only stored secrets as accessed.
Fireworks Ties the Breach to a Group That Resells Stolen Model Access
Fireworks says the evidence points to a financially motivated cybercriminal group whose aim is to sell tokens from unauthorized access to large language models to illicit routing gateways. The practice is known as LLMjacking. Fireworks cited a 2 November publication by Google Threat Intelligence Group and Mandiant describing the technique. It has not named the group. In response, Fireworks revoked and replaced credentials and service accounts, isolated affected systems, and disabled self-serve dataset uploads and downloads for Batch Inference and dataset-based training while the investigation continues.
Notion Says Fireworks Kept Data Covered by a Zero Data Retention Commitment
Notion used Fireworks for classification jobs behind Insights, a feature that shows workspace administrators how their members use Notion AI, and as a general provider for serving open-weight models inside Notion. It also used data volunteered by customers in its Learning and Early Access Program to experiment with training task-specific models on Fireworks. For Insights, Notion sent stored Notion AI requests and responses to Fireworks for analysis. Notion says Fireworks had committed by contract to zero data retention for that use, and that its investigation later found the data had been retained. TechCrunch’s Zack Whittaker reported that Notion confirmed the stolen material was chat transcripts from Notion AI customers who used Insights. According to Notion, a transcript holds any model input or output written to the chat. For most affected workspaces, Notion says, the transcripts carried no metadata tying them to a specific user, workspace or company.
Notion Cut Off Fireworks and Told Customers to Rotate Pasted Credentials
As of 3 November, Notion had disconnected its systems from Fireworks, deleted the data held there, stopped all traffic to the provider and halted training runs. Insights is disabled during the investigation, and Notion no longer serves open-weight models through Fireworks. Owners of affected workspaces with a verified domain can download the exposed text, exactly as it was accessed, from a Security incident review panel in the workspace sidebar, and affected members can view their own records. Notion advises anyone who pasted passwords, API keys or other credentials into Notion AI during the period to rotate them right away. It says it is now auditing critical vendors, including inference providers and data processors, to verify they follow data minimization and zero data retention policies.
Why This Matters for Engineering Leadership
Notion’s disclosure separates two kinds of AI traffic. Its zero data retention program covers the API requests it sends to model vendors to generate responses. Insights worked differently. It depended on conversations Notion already stored, and it shipped those stored conversations to a vendor for a second round of processing. Notion says its retention commitments were never meant to cover features that require stored data. For this flow, the vendor-side protection Notion describes is a contract clause, and Notion learned during its breach investigation that the clause had not held. A retention promise describes another company’s systems, and a buyer cannot observe it directly. Engineering leaders who route user content to inference providers should inventory every feature that sends stored content to a vendor, keep that list apart from live inference paths, and ask each vendor how retention is enforced on the specific endpoints in use. Notion’s advice to rotate credentials pasted into AI chat applies to any team running an assistant over internal work.
FAQ
What data was exposed in the Fireworks AI breach?
Fireworks says an internal application exposed resource metadata and billing information, plus a subset of incomplete model weights and datasets for some customers. It says payment card numbers are not stored by Fireworks and were not affected.
How many Notion customers were affected by the Fireworks AI breach?
Notion says fewer than 1% of Notion workspaces were affected. It has contacted the affected customers directly.
What Notion data was stolen?
Chat transcripts from Notion AI customers who used the Insights feature, including any model input or output written to the chat. Notion says AI Meeting Notes data was not affected unless it was used as input in a Notion AI chat.
Did Fireworks have a zero data retention agreement with Notion?
Yes. Notion says Fireworks gave it a contractual commitment to zero data retention, and that its investigation found data had been retained from the Fireworks endpoints Notion used.
Who was behind the Fireworks AI attack?
Fireworks says the evidence links the attack to a financially motivated cybercriminal group that sells tokens from unauthorized access to large language models, a practice known as LLMjacking. Fireworks has not named the group.
What should affected Notion users do?
Review the affected data in the Security incident review panel in the workspace sidebar and treat it as exposed. Notion recommends rotating any passwords, API keys or other credentials pasted into Notion AI during the affected period.
