/

Notion Says Fireworks Promised to Keep Nothing. Fireworks Says Its Batch API Never Did

Notion AI chat transcripts from under 1% of workspaces were in data attackers reached at Fireworks AI. Notion says it had a zero-retention contract; Fireworks’ CTO says the Batch API it used was never zero-retention.

By

·

Written with AI assistanceHow we reportCorrectionsReport an error

Rows of black tower servers on metal shelving lit by blue LED strips in a dim server room

Notion says a vendor that had promised to keep none of its data kept some anyway, and attackers found it. In a help-center FAQ on the breach at AI infrastructure company Fireworks AI, Notion says Fireworks gave it “a contractual commitment to zero data retention,” yet “during our investigation we learned that data was retained from the Fireworks endpoints that we used.” The exposed data: Notion AI chat transcripts from fewer than 1% of Notion workspaces. Fireworks’ co-founder and CTO, Dmytro Dzhulgakov, replied on X that the FAQ “doesn’t accurately reflect the facts,” because the Batch API Notion used was never zero-retention. On Nov. 13, Notion added that it had reported third-party API keys found in the stolen chats to the companies that issued them.

Key points

  • Notion says Notion AI chat transcripts used for its Insights feature, from fewer than 1% of workspaces, were in data attackers accessed at Fireworks AI.
  • Notion says Fireworks had contractually committed to zero data retention for Notion’s use, but data was retained from the Fireworks endpoints Notion used.
  • Fireworks’ CTO says that account is inaccurate: the Batch API Notion used stores an uploaded dataset until the customer deletes it, and Fireworks documents that.
  • Notion revoked Notion API keys found in the exposed chats on Nov. 11 and reported third-party keys to their issuers on Nov. 12.
Rows of black tower servers on metal shelving lit by blue LED strips in a dim server room
A server room at hosting company BalticServers in 2013. This is not a Fireworks AI or Notion facility and does not show any system involved in the breach. Photo by BalticServers.com via Wikimedia Commons, CC BY-SA 3.0. Resized for this post.

Last checked: Nov 13, 2026. This is a developing story; both companies say their investigations are ongoing.

What leaked, according to Notion

Notion’s security advisory says the exposed transcripts came from Insights, a Notion feature that shows workspace administrators how their members use Notion AI. To build those dashboards, Notion sent stored Notion AI requests and responses to Fireworks to run classification jobs. Notion also used Fireworks as a general inference provider (the service that actually runs a model and returns an answer) for open-weight models inside its product.

Per the FAQ, a transcript contains “any input to or output from the model that was written to the chat.” Notion says that for the vast majority of affected workspaces the chats were “unattributable,” meaning they weren’t labeled with a user or workspace and didn’t include tool-call results. AI Meeting Notes audio, transcripts and summaries were not affected, Notion says, unless someone pasted them into a Notion AI chat.

That last caveat is the whole problem. A chat with no name attached can still contain a name, a customer list, a contract draft or a password. Notion seems to agree: it tells affected customers to “treat this data as if it were exposed” and to rotate any passwords, API keys or other credentials pasted into Notion AI during the incident window. The company also built tools for it. Owners of affected workspaces with a verified domain can download CSVs of the exposed text “exactly as it was accessed,” plus an annotated version that flags potential sensitive information, and can share a link so members see only their own exposed chats.

Notion’s own keys were apparently in there too. On Nov. 11, it says, it revoked exposed Notion API keys found in affected user messages. On Nov. 12 it reported exposed third-party API keys to the vendors that issued them, an update posted to the advisory on Nov. 13. Notion has disabled Insights and stopped serving open-weight models through Fireworks.

The timeline, stitched from both companies

Neither company publishes the whole sequence in one place. Put their bulletins side by side and it looks like this:

  • Oct. 6: Fireworks identifies unauthorized use of internal credentials. Its first bulletin says environment variables, including API keys and tokens stored as secrets for evaluation jobs, were captured in some customer accounts, and adds: “Fireworks does not store prompts or outputs by default, as our platform is ZDR.”
  • Oct. 23–24: Per Fireworks’ Nov. 9 update, a third-party cloud provider hosting Fireworks R&D machines has its own incident on Oct. 23. On Oct. 24 an unauthorized party uses it to reach Fireworks development machines and run a workload that “led to additional customer data impact.” Fireworks says it detected that the same day and has seen no further successful access to its customer platform since Oct. 25.
  • Nov. 2: Fireworks tells Notion a small set of data was accessed. Notion says that set held only synthetic Notion AI testing data.
  • Nov. 3–4: Notion disconnects from Fireworks, then deletes all its data on Fireworks’ systems.
  • Nov. 6: Fireworks tells Notion potential customer data may have been accessed.
  • Nov. 8: Fireworks confirms the final affected dataset, which Notion finds includes Insights chat transcripts.
  • Nov. 9: Notion contacts affected customers; Fireworks publishes its widened bulletin.
  • Nov. 11–12: Notion revokes exposed Notion API keys, then reports third-party keys to their issuers.

Two things stand out. First, Notion’s timeline starts on Nov. 2, nine days after the Oct. 24 intrusion Fireworks describes, and customer data was confirmed on Nov. 8, 15 days after it. Second, Notion’s Nov. 4 deletion was the right move but couldn’t undo this breach: Fireworks ties the newly found customer-data impact to the Oct. 24 workload, more than a week earlier. Deleting the copy shrinks future risk; it doesn’t un-steal anything. Neither company has said exactly when the Notion transcripts were taken.

Three different “zero retention” promises

“Zero data retention,” or ZDR, means an AI provider processes your prompt and answer and then keeps nothing. This incident shows the term covers at least three different promises, and it’s easy to hear the wrong one.

Notion’s promise to its customers. Notion’s FAQ says its enterprise ZDR commitment covers the API requests it sends to third-party model vendors to answer Notion AI questions. It is “not intended to cover” features where Notion has to store the data to deliver them, such as chat history and Insights. So an enterprise customer with ZDR in its contract could still have had chats in this dataset, because Insights ran on stored copies.

Fireworks’ promise to Notion. For the Insights analysis, Notion says it relied on a separate contractual ZDR commitment from Fireworks, and that data was retained anyway. We haven’t seen that contract, and neither company has published it.

Fireworks’ product setting. Fireworks’ documentation says most of its platform never stores prompts or generations, but “a few features do store customer content, such as stored responses, batch jobs, and training.” Its account-wide Zero Data Retention policy, an Enterprise feature that is “Off by default,” blocks those features by rejecting batch inference jobs, dataset uploads, evaluation jobs and more. Turning it on “doesn’t delete anything already stored on the account.”

Dzhulgakov’s Nov. 12 post leans on exactly that page. Fireworks’ Inference API is ZDR and wasn’t affected, he wrote, but the “Batch API (used here) is not ZDR because it requires uploading a dataset first, before starting a batch job later.” Until the customer deletes it, he says, the dataset stays on the platform and can be downloaded. CEO Lin Qiao’s Nov. 9 post, as quoted by VentureBeat, likewise said no impact was detected to Inference “(excluding batch inference).” Fireworks’ Nov. 9 bulletin also says it has disabled self-serve dataset uploads and downloads used by Batch Inference while it investigates.

Our read: Both statements can be true at once, which is the uncomfortable part. A vendor can sign a no-retention clause, ship a batch product that stores files by design, and leave the setting that would block it switched off. Whether the contract covered batch jobs, whether Notion should have known its classification jobs ran through stored datasets, and who should have deleted them are questions only the contract and the two companies can answer. So far they disagree in public.

Notion’s volunteers and a second Fireworks breach

Notion’s FAQ adds a detail that hasn’t drawn much notice. Members of its Learning and Early Access Program (LEAP) “volunteer their data to help Notion test new AI technology,” and Notion says it used that volunteered data in the past “to experiment with training task-specific models on Fireworks.” Notion doesn’t say whether any LEAP data was in the affected dataset, so we don’t know. But it confirms Notion data sat on Fireworks for training work as well as inference.

Fireworks presents all this as an update to its October incident. Bloomberg Law called it the company’s second known hack in a matter of weeks, and TechCrunch the second breach in as many months. Fireworks says it notified law enforcement and brought in Palo Alto Networks’ Unit 42 for forensics. Its Nov. 9 update lists what the affected internal application exposed: resource metadata, billing details such as billing and shipping addresses, invoice and payment history and credit usage, and, for some customers, “a subset of incomplete model weights and datasets.” Fireworks says it doesn’t store payment card numbers. A Fireworks spokesperson told Bloomberg Law the company is working with customers “including Notion, Cursor and others.”

Why the attackers wanted AI access

Fireworks says the evidence links the intrusion to a financially motivated group trying to sell stolen access to large language models through illicit routing gateways, a scheme known as LLMjacking. Google’s Threat Intelligence Group and Mandiant described the trend on Nov. 2: in incidents they handled, fraudulent model charges topped $100,000 in a single day, and cumulative charges at some organizations reached $1 million before access was cut off. Google calls the resale middlemen “LLM Mixers.”

That framing matters for Notion users. If the goal was resold model access, the chat transcripts may have been collateral rather than the target. But stolen API keys sitting inside those chats are exactly the kind of thing LLMjacking crews monetize, which is why the key revocations and vendor notices are the most practical part of Notion’s response. It’s the same economics behind the stolen AI keys Railway said ran up about $740,000 this week.

Who it affects

If you use Notion and neither you nor your workspace owner got a notice, Notion says no action is needed right now. If you did get one, the FAQ’s advice is direct: review the CSVs, assume the content is exposed and rotate any credential that appears in it. Admins without a verified domain can’t download the affected-content files, per Notion’s FAQ, so domain verification is the first step.

For companies buying AI services, the lesson is narrower than “don’t trust vendors.” Ask which features store data (batch jobs, fine-tuning, evaluations, stored responses), ask whether the vendor’s own retention setting is on for your account, and ask who deletes what and when. Notion says it is now auditing critical vendors such as inference providers to “verify” they follow data-minimization practices, including ZDR. Verification is the right word, and it’s coming after the fact.

What to watch

  • The contract question. Whether Notion or Fireworks explains what the ZDR clause covered, and whether it named batch jobs or stored datasets.
  • LEAP data. Whether Notion says any volunteered LEAP data was part of the affected Fireworks dataset.
  • Other Fireworks customers. Whether Cursor, Harvey or other customers disclose affected data, and whether more of Fireworks’ customers find stored datasets in the leak.
  • The cloud provider. Fireworks still hasn’t named the third-party provider breached on Oct. 23, or said whether that company’s other customers were hit.
  • Key abuse. Whether any third-party key that Notion reported is tied to fraudulent model usage.

Hype in. Signal out.

Image: server room at BalticServers, 2013, by BalticServers.com via Wikimedia Commons, licensed CC BY-SA 3.0; resized. Not a Fireworks AI or Notion facility.