The Fireworks AI breach keeps widening: Notion has confirmed that attackers at AI infrastructure provider Fireworks AI obtained Notion AI chat transcripts. According to Notion, fewer than one percent of workspaces are affected. Its own systems were not compromised; the data sat with the vendor. In its latest update of 13 November 2026, Notion says it revoked API keys found in the exposed messages or reported them to the vendors that issued them. Fireworks itself had already acknowledged that billing information, datasets and parts of model weights were affected in addition to credentials.
Fireworks AI breach: what happened at Fireworks
Fireworks AI runs a platform that companies use to serve, fine-tune and evaluate open AI models. The company says it identified unauthorized use of internal credentials on 6 October. Its first security bulletin of 9 October said a third party had accessed environment variables used in evaluation jobs in a limited number of customer accounts, including API keys and tokens stored as secrets. At the time, Fireworks said it had found no evidence of access to customer models, datasets, training data or inference traffic.
That picture has since changed. In its update of 9 November, Fireworks writes that a third-party cloud provider hosting Fireworks R&D machines suffered its own security incident on 23 October. On 24 October, the attacker used it to reach Fireworks development machines and run an unauthorized workload there. Fireworks says it detected this the same day. Since 25 October, it has found no further evidence of successful access to the customer platform.
According to Fireworks, the affected internal application gave access to the following data:
- Resource metadata such as identifiers, display names, descriptions and creation times
- Billing information such as billing and shipping addresses, invoice and payment history, account balance and credit usage. Fireworks says it does not store payment card numbers.
- Model weights and datasets: for certain customers, a subset of incomplete model weights and datasets
Fireworks notified law enforcement and is working with Palo Alto Networks Unit 42 on forensic analysis. The company says the evidence points to a financially motivated criminal group that resells unauthorized access to large language models, a practice known as “LLMjacking”.
Fireworks lists four areas of countermeasures: credentials and service accounts were revoked and replaced, unnecessary elevated privileges removed and the creation of new credentials more tightly controlled. Network pathways were restricted and affected systems isolated. Access to data storage and analytics systems was tightened, and logging and monitoring were strengthened. Fireworks says the investigation is ongoing and may reveal additional information.
Which Notion data was exposed
Notion used Fireworks in two ways: for classification jobs behind its “Insights” feature, which helps customers understand their own AI usage, and as a general inference provider for open-weight models inside the product. According to Notion’s security advisory, the incident unfolded as follows:
- 2 November: Fireworks informs Notion that a small dataset was accessed. It reportedly contained only synthetic test data.
- 3 November: Notion disconnects all systems from Fireworks.
- 4 November: Notion deletes all of its data in the Fireworks system.
- 6 November: Fireworks reports that customer data may also have been accessed.
- 8 November: Fireworks confirms the final affected dataset.
- 9 November: Notion contacts affected customers.
- 11 November: Notion revokes Notion API keys found in affected messages.
- 12 November: Notion reports third-party API keys found in affected messages to the vendors that issued them.
The exposed data consists of Notion AI chat transcripts used for Insights. For the vast majority of affected workspaces, Notion says they are “unattributable”: the chats were not labeled with an author or workspace and did not contain tool call results. The content itself can still reveal a lot, for example when users pasted names, customer data or credentials into the chat. That is exactly what Notion’s recommendations address. No other Notion features are affected so far, and the investigation is ongoing.
Notion has disabled Insights and no longer serves open-weight models through Fireworks. Affected customers can review the exposed data inside their workspace, following instructions sent by email.
What affected teams should do now
Notion recommends treating the affected data as if it had been made public. Anyone who pasted passwords, API keys or other credentials into Notion AI during the relevant period should rotate them. In practice:
- Check whether you received a notice from Notion. Without one, Notion has not identified your workspace as affected so far.
- Review the data provided in the workspace for credentials, personal data and confidential business information.
- Revoke and reissue any keys and passwords found with the respective provider, even if Notion has already passed on a report.
- If personal data is involved, check with your data protection officer whether notification or disclosure obligations apply.
Fireworks customers who received their own notice should, according to Fireworks, replace revoked Fireworks keys, rotate any third-party credentials named in the notice with the issuing provider and review those services’ access logs. Simply removing a stored secret does not invalidate the underlying credential.
Zero data retention does not cover everything
The case exposes a gap that many companies underestimate with AI vendors. Fireworks says its Inference, Training SDK and Nexus products use zero data retention by default and were not affected. Other features do store data, however, such as datasets for batch inference and training. That is where the exposed information sat. Fireworks has temporarily disabled self-serve uploads and downloads of such datasets.
For companies building AI features into their products, this raises a few questions worth checking. Which subprocessors handle prompts and responses? Which features store data persistently, even when the vendor advertises zero data retention in general? And how quickly do you learn about an incident at the vendor? In Notion’s case, six days passed between the first notice and confirmation that customer data was involved.
Earlier cases such as the OpenCode breach involving stolen AI API keys and the LiteLLM supply chain attack show that AI credentials are a lucrative target. Teams using Notion can find setup basics in our Notion setup guide for teams. The most important rule applies regardless of vendor: credentials do not belong in AI chats.




