All posts
2 min readby Romiel Inolino

Some Notion AI chats were exposed through a breach at Notion's AI vendor Fireworks. Know which vendors see your automation data

AI securityNotiondata privacyvendor riskAI automation

Notion was not breached, but some of its customers' AI chats were, because the data passed through a vendor most users had never heard of.

What happened

Notion says one of its vendors, Fireworks AI, notified it of a security incident affecting fewer than 1% of Notion workspaces. Notion has contacted affected customers and added a Security incident review page where workspace owners can download the affected content as CSV files. Its advice is to treat that data as exposed, and to rotate any passwords, API keys or other credentials people pasted into Notion AI during the period.

The Register reports that Fireworks, which lets customers train and run AI models, disclosed a second attack: a third party cloud provider hosting its research and development machines was compromised on October 23, and attackers used that access the next day to reach Fireworks development systems. The exposed data includes resource metadata, billing records and, for certain customers, a subset of incomplete model weights and datasets. Fireworks first disclosed an earlier incident on October 9, where attackers accessed environment variables that included some customers' API keys and tokens.

Per The Register, Notion told affected customers the exposed transcripts had no separate information tying them to a workspace or user, though Reddit users pointed out the conversations themselves can contain identifying details.

My take

This is the part of AI automation most teams never map. Your client trusts the app they signed up for. That app calls a model host, which runs on a cloud provider, which may log or store things you never see.

Three things I now do on every build:

  1. List every vendor a record passes through, including the model provider behind each AI step, not just the tools you log into.
  2. Keep secrets out of prompts. API keys belong in a credential store, never in a chat, a Notion page or a workflow note that gets sent to a model.
  3. Strip or mask personal data before it reaches an AI step that does not need it.

If a vendor two steps away gets breached, you want a short list of what was exposed, not a guess.

More posts