Notion has confirmed that AI chat transcripts from some of its workspaces were exposed in the Fireworks AI breach, and the transcripts are the kind of data that is hard to take back once it is out. Fewer than 1% of Notion workspaces may have been affected, according to Notion’s security advisory, and the exposed material is chat transcripts from the Insights feature, the tool that helps customers understand their AI usage. Notion describes the transcripts as unattributable for the vast majority of affected workspaces, with no author or workspace labels and no tool call results attached, but the conversations themselves could contain identifying details, as Reddit users pointed out in a discussion reported by The Register.
The Notion confirmation is the human face of a wider disclosure. On 9 November, Fireworks AI told customers that its October security incident had reached further than it first disclosed: billing information, and for a limited set of accounts, secrets, datasets and incomplete model weights. It was the second break-in at the Nvidia-backed AI infrastructure provider in a matter of weeks, according to Bloomberg Law, and the shape of the month is this: the breach did not happen once. It happened twice, 18 days apart, and the fuller picture took 34 days to surface.
What Notion confirmed, and when
Notion’s advisory lays out a week in which the picture changed three times.
| Date | What happened |
|---|---|
| 2 Nov, evening | Fireworks tells Notion a small dataset was accessed; Notion’s first review finds only synthetic testing data |
| 3 Nov, morning | Notion disconnects from Fireworks, deletes all data held in the Fireworks system, and stops all communication with Fireworks |
| 6 Nov | Fireworks warns malicious actors may have accessed potential customer data; Notion opens a second investigation |
| 8 Nov | Fireworks confirms a new dataset; Notion’s analysis finds it contains the Insights chat transcripts |
| 9 Nov | Notion contacts affected customers |
The transcripts came from the Insights feature, which uses Fireworks for classification jobs that help customers understand their AI usage. Fireworks also served as a general inference provider for open-weight models inside Notion. Both are now off: Notion has disabled Insights during the investigation and is no longer serving open-weight models through Fireworks, according to the advisory. One customer has already gone public about being potentially affected. Render founder and chief executive Anurag Goel shared a Notion notification saying his workspace was among the fewer than 1% that may have been affected, The Register reported. Notion stresses that this was not a breach of its own systems, and that no other Notion features were affected.
Two details come from other outlets, not from the advisory itself. Notebookcheck reported that on 11 November, Notion revoked Notion API keys it found in affected messages during its investigation. And BleepingComputer reported that Notion has published a FAQ page noting Fireworks had contractually committed to zero data retention for Notion data. Neither detail appears in the advisory as currently published, so both stay attributed to the outlet that reported them.
Two break-ins, 18 days apart
The first break-in came to light on 6 October, when Fireworks identified unauthorized use of its internal credentials. An unauthorized third party accessed environment variables used in evaluation jobs in a limited number of customer accounts, and Fireworks believes API keys and tokens stored as secrets were captured in certain affected accounts, according to its original bulletin. Notifications to identified affected customers began on 7 October. The first bulletin went out on 9 October, employee API keys were rotated the same day, and on 10 October Fireworks revoked additional user API keys and rotated service-account tokens.
The second entry came through the side door. On 23 October, a third-party cloud provider hosting Fireworks research and development machines had its own security incident, and on 24 October the unauthorized third party used it to reach Fireworks development machines and run an unauthorized workload that led to additional customer data impact, according to the 9 November update. Fireworks detected the activity the same day and took immediate remediation measures, and as of 25 October it had not detected further evidence of successful unauthorized access on the customer platform. The provider has not been named in any source read so far.
Here is the part that matters most, and it is our own reading of the two bulletins. The 9 October bulletin said that, based on its investigation at that time, Fireworks had found no evidence of unauthorized access to customer models, datasets, training data, personal information, or inference traffic within its security perimeter. The 9 November update expanded those findings: it added impacts to account and billing information, and for a limited set of accounts, to datasets and model weights. Thirty-four days passed between the first detection on 6 October and the wider disclosure on 9 November, and the two entries were 18 days apart. Wiz’s Black October report lists Fireworks first among the AI developer platforms that disclosed October incidents, with 6 October as the date Fireworks identified its incident, and Fireworks has since made a second, wider disclosure on 9 November.
Fireworks chief executive Lin Qiao has pushed back on one fear about the incident. Responding to concerns raised by Cantina Security chief executive Hari Mulackal, Qiao said, “From what we see attackers are not after weights” and “Most downloads were partial and stopped,” according to The Register, which cited Qiao’s posts on X. That is the company’s position, not an independent finding, and it sits next to the confirmed fact that incomplete model weights were accessed for a limited set of accounts.
What Fireworks says was taken, and what was not
| What was accessed | What was not |
|---|---|
| Billing and shipping addresses, invoice and payment history, account balance and credit usage | Payment card numbers (Fireworks says it does not store them) |
| Resource metadata: resource identifiers, display names, descriptions, creation times | Inference, Training SDK and Nexus (zero data retention by default; no impact detected so far) |
| Secrets, datasets and incomplete model weights (a limited set of accounts) | Not applicable |
On motive, Fireworks says the evidence points to a financially motivated cybercriminal group aiming to sell tokens from unauthorized access to large language models to illicit routing gateways, activity known as LLMJacking. Google’s Threat Intelligence Group and Mandiant published a blog post about the rise of LLMJacking on 2 November. No group has been named, and the LLMJacking link is Fireworks’ assessment, not an independently confirmed attribution. Whether the accessed data was used, sold or leaked has not been disclosed in any source read.
The response so far has two names on it. Fireworks notified law enforcement and brought in Palo Alto Networks’ Unit 42 for forensic analysis and security hardening, and it has revoked or replaced credentials and service accounts, removed unnecessary elevated privileges, isolated networks, tightened access to data storage and analytics, and secured its logging and monitoring. It has also disabled self-serve dataset uploads and downloads used by Batch Inference and dataset-based training while it investigates. Bloomberg Law described the breach as involving tens of thousands of files containing customer data. Separately, Fireworks lists over 10,000 customers on its website, including legal tech firm Harvey, productivity software startup Notion and coding company Cursor. No confirmed total of affected customers has been published, and Cursor’s exposure, if any, remains unconfirmed.
What to do if you use Notion AI
If you use Notion AI, check your email for a notice from Notion, open the impacted data inside your workspace, and treat the data as exposed. Rotate every password, API key or credential you pasted into Notion AI, and review the transcripts for secrets you forgot were there. The text can still identify you.
- Open Notion’s email and follow its instructions to review the impacted data inside your workspace.
- Treat the exposed chats as exposed data, even though Notion’s advisory says the transcripts were unattributable.
- Rotate any password, API key or credential you pasted into Notion AI during the incident.
- Read the impacted transcripts and look for secrets, addresses or account details you forgot were there.
- Replace the Fireworks credentials listed in your incident notice, and rotate the third-party credentials it identified.
- Watch for phishing that quotes details from your chats, and report anything suspicious to Notion.
Steps 1 to 3 follow Notion’s advisory. Steps 4 to 6 are standard security practice, not instructions from Notion or Fireworks.
What to watch next
Five questions are still open. Whether Cursor’s data was exposed (Bloomberg Law reported that Fireworks is working with Cursor, among others, to manage the incident), and whether Fireworks names the third-party cloud provider behind the second entry. Whether a confirmed count of affected customers appears, and whether any of the accessed data is used, sold or leaked. Whether the other platforms in Wiz’s original nine-platform list report wider impact of their own, and whether Notion’s advisory changes again. Fireworks says its investigation is ongoing and may reveal additional information.
Wiz found no single vulnerability or dependency linking the nine incidents in its Black October list, and the sources do not establish one common actor behind all nine. That list includes Railway, whose own breach we covered separately.


