Notion's AI Vendor Promised to Keep Nothing. A Hack Found Notion's Chat Transcripts Anyway.

Notion says its AI vendor promised to keep nothing. The vendor says the endpoint Notion used was never covered. Here is how to check what "zero retention" covers.

Notion says a hack at Fireworks AI, a company that runs AI models for it, exposed Notion AI chat transcripts from fewer than 1% of Notion workspaces. The transcripts fed a Notion feature called Insights. For most affected workspaces they carried no author or workspace label, but Notion still tells those customers to treat the data as exposed and to change any password, API key or other credential they pasted into Notion AI.

The hosts are AI characters.

Listen to this story
2:22

The part every company buying AI should read is the promise. Notion says Fireworks gave it a contract for zero data retention, meaning nothing it sent would be kept, and that Notion learned during the investigation that data was retained from the Fireworks endpoints it used. Fireworks' co-founder and CTO, Dmytro Dzhulgakov, replied on X that Notion's account is inaccurate: the Batch API used here is not zero-retention, he says, because a dataset is uploaded first and stays downloadable until the customer deletes it. The two companies still disagree.

Notion is not the only Fireworks customer counting the damage. On November 13, Cursor said its own October breach "started at Fireworks": an access key Cursor had issued to Fireworks was exposed, and the attacker used it to copy coding-agent traces, mostly from 2025, from sessions not in Cursor's Privacy Mode. A few of those traces held credentials nobody had scrubbed, and that is how the attacker got further into Cursor. Fireworks' bulletin, last updated November 9, does not mention Cursor.

Fireworks' own docs show why both can sound right. Zero retention there is an Enterprise account setting, and it works by refusing the features that store customer content: stored responses, batch jobs and training. Fireworks says the products that are zero-retention by default showed no detected impact. A contract can say "zero retention" while the endpoint in use keeps data, because the promise attaches to endpoints and settings, not to the company.

Notion's own fine print works the same way. Its help page says its zero-retention commitment to enterprise customers is not meant to cover features that store requests and answers to work, such as chat history and Insights.

Jon's agents send work to many models through OpenRouter, a router that offers the same choice request by request.

Check what "zero retention" covers

  • Ask the vendor which exact endpoints are zero-retention, and whether it is on for your account or only available. Treat batch jobs, uploaded datasets, stored responses and fine-tuning as kept unless the vendor says in writing they aren't.
  • Check your own product's features. Chat history, saved threads and analytics like Notion's Insights store data by design, whatever the model vendor promises.
  • On OpenRouter, add "provider": {"zdr": true} to a request and it routes to zero-retention endpoints, or turn it on for the whole account in privacy settings. Unless the account says otherwise, self-serve accounts also allow "Safety Retention" endpoints, which may keep data for safety review; exclude them in the same settings.
  • See which models survive the switch first. OpenRouter keeps the eligible endpoints in a live list, updated when a provider's policy changes; a model with no endpoint there won't route under the setting. On November 14 it held 1,108 endpoints, including Jev 1.14, Claude Haiku 5.5, Drex Max and Grok 4.7, but still not Google's new Gemini Decision 1. The list also includes the Safety Retention endpoints, so check those against the account setting.
  • Keep credentials out of prompts. Notion's first instruction to affected customers was to rotate anything pasted into its AI.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.