The breach did not start at Fireworks AI. That is the part worth understanding.

Fireworks, a platform developers use to run open models, disclosed that an attacker got in through a third-party cloud provider hosting its research and development machines. The intrusion at that provider happened on October 23. By October 24 the attacker had reached Fireworks development machines and was running an unauthorized workload. Fireworks says it detected the activity the same day.

What was taken, according to Fireworks' advisory: customer billing information, some datasets, and for some customers partial or incomplete model weights. Payment card numbers were not affected because Fireworks says it does not store them.

The downstream casualty is more instructive. Notion confirmed that the incident exposed chat transcripts from Notion AI customers who used the Insights feature, a tool that shows workspace administrators how members use AI. Those transcripts are the material Notion was sending to Fireworks to power the feature, and they contain, in Notion's own words, any input to or output from the model that was written to the chat.

About 1% of Notion workspaces were affected. The timeline was slow by security-incident standards: Fireworks alerted Notion on November 2 about a dataset it believed was synthetic test data. Notion disconnected its systems from Fireworks on November 3 and deleted its data there. On November 8 Fireworks confirmed the dataset included real customer chats. Notion began contacting affected customers on November 9, and on November 11 it revoked API keys that customers had pasted into affected messages during the investigation.

Notion also disclosed something that will worry every team with a vendor contract. Fireworks had given it a contractual commitment to zero data retention. During the investigation, Notion says it learned data was in fact retained on the endpoints it used. Notion's own explanation is that its zero-retention commitments cover API requests to model vendors, but not product features that need stored data, such as chat history and Insights.

The attackers' motive appears to be financial rather than espionage. Fireworks says the evidence points to a group selling unauthorized access to large language models, the practice known as LLMjacking, where stolen credentials are resold so buyers can burn someone else's inference quota. This was the second Fireworks incident in two months, after an earlier security event in October.

Unattributable is not the same as harmless. Notion says most of the exposed transcripts could not be traced to a specific workspace or author, and included no tool-call results. But a transcript still contains whatever a person typed: a password, an API key, a customer list, a draft contract. The absence of a label on the data does not reduce what the data reveals.

The practical lesson for teams routing prompts through third-party inference: your chat logs are exactly as safe as your vendor's least-audited cloud supplier, and your data-retention promise is only as good as the vendor's storage configuration. Notion is now auditing inference providers and data processors for data minimization. Every company that ships an AI feature should be doing the same audit before the incident, not after it, and should assume that any prompt containing a secret has already been logged somewhere it does not control.