// ai security

Notion Fireworks AI Incident: Zero Data Retention Failed

Notion workspace sidebar opened to the question-mark menu and Security incident review, the desktop path owners use to download affected-member and content CSVs after the Fireworks AI incident.

The Notion Fireworks AI incident began when Fireworks AI notified Notion of a security incident that Notion says affects fewer than 1% of workspaces. It is a processor-control failure: a zero data retention commitment that did not match what the endpoints actually kept.

Key takeaways

  • Fireworks AI notified Notion of an incident Notion says affects fewer than 1% of workspaces, and the investigation is still open.
  • Notion used Fireworks for Insights classification and open-weight inference, and it cut the connection as of November 3.
  • Fireworks had contractually committed to zero data retention; during its investigation Notion found data retained on the endpoints it used.
  • Treat affected chats as exposed and rotate pasted credentials; content CSVs require a verified domain and a desktop owner session.

What is confirmed in the Notion Fireworks AI incident?

Notion confirmed that Fireworks AI notified it of a security incident affecting fewer than 1% of Notion workspaces. It has contacted the customers it identified, and the investigation continues. The notice is titled “Notice about the Fireworks AI incident that may impact some Notion workspaces.”

The notice does not describe how the incident occurred, the exact exposure window, or the volume of data. If neither a user nor the workspace owner has heard from Notion, the workspace is not identified as affected, and Notion says no action is needed right now. The notice gives no date for when Fireworks told Notion, which is the date a regulatory notification clock would run from.

Which product paths sent Notion data to Fireworks?

Two product paths sent Notion data to Fireworks: classification jobs behind Insights, and inference for open-weight models served inside the product. Insights was how administrators were meant to see their own AI usage. Notion sent stored Notion AI requests and responses, Fireworks ran the analysis, and Notion visualized the result.

Open-weight means the model weights are publicly downloadable. Prompts sent to an externally hosted open-weight model still leave the customer environment and follow that provider’s processing and retention. A menu option labeled open-weight can still mean a third party hosts the call.

Notion has disabled Insights during the investigation and has stopped serving open-weight models through Fireworks. A chat transcript means any model input or output saved as part of the conversation. AI Meeting Notes audio, transcripts, and summaries were not affected unless that material was used as an input in a Notion AI chat.

Learning and Early Access Program (LEAP) customers volunteer data so Notion can test new AI technology. In the past, Notion used that volunteered data to experiment with training task-specific models on Fireworks. The notice does not place LEAP in the affected set and does not say those records were retained.

Stripping user and workspace IDs can still leave a company name, a person’s name, a domain, source code, or customer details inside the conversation.

Why did Fireworks' zero data retention promise fail?

The zero data retention commitment failed because Notion learned, during its investigation, that data was retained from the Fireworks endpoints it used. Notion says this directly:

Unfortunately, during our investigation we learned that data was retained from the Fireworks endpoints that we used.

Fireworks had given Notion a contractual commitment to zero data retention, and the endpoints Notion called kept data anyway.

Notion’s own zero data retention statement covers API requests sent to third-party model vendors used to generate responses to Notion AI requests. Notion says it works with each vendor so those API endpoints are zero data retention compliant. The expectation is that the vendor does not retain the request and response associated with that API call. The commitment is not intended to cover features that require Notion to store requests and responses. Notion names chat history and Insights.

Insights is where the two retention regimes collide. Notion stored the chats for history and admin analytics, which its own zero data retention terms leave out, then sent those stored requests and responses to Fireworks. Notion says it applied data-minimization practices to that analysis. Fireworks’ contract said that copy would not be kept, and that is the copy Notion says Fireworks retained.

GDPR Article 28 requires controllers to use processors that provide sufficient guarantees and to govern the processing by contract. A zero-retention sentence does not prove that the live endpoint applied the setting. The clause, the endpoint identifier, and a log sample are three different artifacts. Only the log sample shows retention.

As of November 3, Notion says it disconnected its systems from Fireworks, deleted all data in the Fireworks system, stopped sending all data and traffic to Fireworks, and halted all training runs. No exposure start date is stated. Treat the deletion claim as containment, not remediation: the notice does not name backups, observability platforms, support exports, subprocessors, or a training set already copied out.

Notion says it is auditing critical vendors, including inference providers, to verify data-minimization and zero-retention practices. It does not name other inference vendors that still receive Notion prompts, and it does not describe how the Fireworks environment was accessed.

What should security leads treat as exposed?

Notion’s instruction is to treat the data as if it were exposed: rotate any passwords, API keys, or other credentials pasted into Notion AI during the relevant timeframe. The public notice does not pin that timeframe to calendar dates. Bound it from the source-content dates in your export. A retained transcript is evidence of exposure. Issuer authentication and usage logs are what show whether a secret was later used.

Those chats left Notion as a standing copy of prompts and outputs. That path is one of the LLM data exfiltration channels that matter in a vendor-retention failure.

Workspace owners open Security incident review from the workspace sidebar, then the ? menu, in a desktop browser on the latest Notion. The review is unavailable on mobile. Owners can click Download affected members to generate a CSV of names, emails, membership status (active or not), user IDs, and the count of how many rows of affected data relate to each user. Owners use Copy link so each member gets a CSV of only that member’s own affected data.

Content files require a verified domain, and only the workspace owner can see them. Customers without that access are told to contact their account team. affected_content.csv is the workspace data affected by the incident, exactly as it was accessed. affected_content_annotated.csv is the same data plus metadata. In the annotated file, __metadata includes the chat transcript text itself, exactly as it was accessed. That text is what the notice says was exposed. Added context was not exposed: the Notion record (row) ID, the workspace ID, the user ID the content is attributed to, the conversation thread ID, and the date of the source content.

__analysis flags mark potential sensitive information so credentials can be rotated first. Flags are triage. Custom tokens, secrets split across messages, encoded material, private keys, and in-house credential formats slip past generic detectors. Preserve the CSVs, review timestamps, and related audit logs in an access-controlled store, and hash the exports.

How do you test an inference vendor's zero-retention claim?

Test an inference vendor’s zero-retention claim on the exact endpoints that will see production traffic, before those endpoints are allowed to see it. The contract states the rule. Logs, backups, and training jobs show whether the rule is real. A conventional web test would not have exercised vendor retention on those classification and inference endpoints. That is the gap covered in what a web pentest misses in AI components.

Before production prompts move, collect endpoint names and regions, plus a statement that no shadow route is enabled; every log, trace, and analytics sink, with a retention value you can measure; subprocessors and backup copies, including when each backup expires; any training, evaluation, or human-review path that can read the same prompts; a deletion attestation that names each store; and permission for one canary prompt with a written result after the window. A canary string sent on the production route should be gone once the stated window closes.

For the next processor, an AI supply chain security audit asks those questions before the first production prompt. LLM application penetration testing checks what your own product does with the answer, including chat history and admin analytics that ship stored prompts back out.

FAQ

Was my Notion workspace affected by the Fireworks AI incident?

Notion says the incident affects fewer than 1% of workspaces and that it has contacted the customers it identified. If neither you nor the workspace owner has heard from Notion, Notion says the workspace is not identified as affected and no action is needed right now. The investigation is still open.

What Notion AI data did Fireworks retain?

Notion says it learned, during its investigation, that data was retained from the Fireworks endpoints it used, after a contractual zero data retention commitment. For Insights, Notion had sent stored Notion AI requests and responses to Fireworks. The notice does not publish the volume, the exposure window, or the access path.

Which Notion features used Fireworks AI?

Notion used Fireworks for classification jobs behind Insights and as an inference provider for open-weight models. Insights is disabled, open-weight serving through Fireworks has stopped, and the disconnect, deletion, traffic stop, and halt of training runs are dated November 3. Meeting Notes were out of scope unless used as Notion AI chat input.

Should I rotate API keys pasted into Notion AI?

Yes, if the workspace is affected and those secrets were pasted into Notion AI during the relevant timeframe. Notion says to treat the data as exposed and to rotate passwords, API keys, and other credentials. Bound the timeframe from dates in your export, and confirm later use in the issuer's authentication logs.

If classification jobs or open-weight inference already send your prompts off tenant, agent::offense can test that processor before the next retention gap: start with an AI supply chain security audit.


// get started

Work with AgentOffense

Tell us about your target and goals. We’ll reply with scope and a fixed-price quote — usually within one business day.

./request_engagement