The Fireworks AI breach: what leaked from Notion AI, and how to keep your AI chats out of the next one
Your chats are never on our servers, because we never receive them. With Outlier's local models and web search off, your words stay on your Mac. The existing lsof capture below shows what we measured; you can run the same check yourself.
What happened in November 2026
A breach at Fireworks AI, an AI infrastructure provider used by Notion, exposed Notion AI chat transcripts from a small number of workspaces. The transcripts came from the dataset for Notion's Insights feature, which is now disabled. BleepingComputer reported this on 11 November 2026. BleepingComputer, 11 November 2026. Source read 2026-11-20 by our reviewer.
The same report says Notion described the transcripts as unattributable for the vast majority of affected workspaces. They were not labelled with who wrote them or which workspace they belonged to. That limits what the records reveal about their authors; it does not mean the text itself was absent.
Fireworks said an attacker reached its development environment through a third-party cloud provider on 23 October. An earlier October intrusion reached API keys and tokens stored as secrets for a limited number of accounts. Notion cut traffic to Fireworks and deleted its data by 3 November. Fireworks had contractually committed to zero data retention. These details come from the 11 November BleepingComputer report, read 2026-11-20 by our reviewer.
The incident also affected Uber driver information. Gizmodo reported on 13 November 2026 that Uber confirmed Fireworks AI was the provider. State notices put the number at at least 7,870 people across three states. Gizmodo, 13 November 2026. Source read 2026-11-20 by our reviewer. This was a separate kind of data from the Notion transcripts.
Why the provider behind the app matters
You may choose one chat app while another company runs the model behind it. A chat sent to that service can pass through more than the company whose name is on the screen. The Fireworks event shows why asking only about the app's own servers leaves part of the question unanswered.
A retention promise is useful to read. Ask which datasets and features it covers, which other providers handle the text, and how deletion works. A promise about storage and a design that never sends the chat to a provider address different parts of that question.
What you can do today
- If you pasted an API key into a cloud chat, rotate it with the service that issued it. Notion advised affected users to rotate any passwords, API keys or other credentials entered into affected chats, as described in the 11 November report, read 2026-11-20 by our reviewer. Removing a chat does not change the key.
- Check which AI vendors your tool uses. Read its provider list and ask your workspace administrator which features are enabled.
- Choose where to do the work before pasting private text. Use a local model when the task should stay on your Mac.
Where a local chat fits, with a check you can repeat
Outlier runs local model inference on your Mac. We never receive those chats, so they are never stored on our servers. Optional web search sends a search query out, and choosing a cloud model with your own API key sends that chat to the provider you choose. Keep both off for local work.
Existing home page measurement: Captured 2026-11-04 on Outlier 1.11.992 while Core was answering, web search off. A past capture, not a live reading.
It was sampled 22 times across a 24s answer. Sockets leaving this Mac: 0. Every socket the engine held used 127.0.0.1. See the original lsof capture.
While a local answer is running, open Terminal and run:
lsof -a -p $(pgrep -x outlier-cli) -i -n -P
Keep -a: it joins the process and network filters. Without it, lsof combines them with OR and includes other processes' sockets. The Mac offline check explains the output and its limits. This is a sample of one engine, not a test of every process or every moment. Local files still need care, including backups and software already on the Mac.
Download Outlier to try a local model. Read the full network-use list before choosing optional features.