AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Cyberattack on Fireworks AI allowed hackers to access Notion customers’ chat transcripts

Collected Nov 12, 2026

Fireworks AI, a platform developers use to reach open AI models, has confirmed that attackers compromised some of its customers' data. According to the company's security advisory, the intruders went through its cloud provider to reach a Fireworks development system, and from there obtained customer billing information and some datasets. The disclosure follows a separate security incident at the company in October, making this the second breach affecting Fireworks in two months. Fireworks says it believes the attackers were financially motivated and that their objective was selling access to large language models, a scheme in which unauthorized third parties consume model usage without spending their own resources.

Notion, the online document editor and notetaker, is one confirmed affected customer. It said the incident touched roughly 1% of its own customers. According to Notion, Fireworks alerted it to a security incident and later confirmed that hackers had taken some Notion customer data. On November 3, Notion disconnected its systems from Fireworks, deleted the data held in those systems, and cut all connectivity between the two. This week Notion confirmed that what was taken were chat transcripts belonging to Notion AI customers who used the Insights feature. Notion said the transcripts capture whatever was written to the chat, including any input to or output from the model. It added that much of the data could not be attributed, so it could not be determined which customer a given record came from.

Fireworks describes its customer base as tens of thousands of organizations worldwide, including the legal tech company Harvey and the AI company Cursor. Neither of those named customers is identified as affected in the disclosure; Notion is the one confirmed case.

For developers, the mechanics matter more than the headline. Fireworks sits in the inference layer: teams point their applications at its API to run open-weight models rather than hosting and scaling those models themselves. That arrangement concentrates risk. A single development system inside the provider can hold credentials, logs, request traces and billing records that span many downstream products, and Notion's case shows how far that reach extends. The stolen material was not Notion's own database but conversational content generated when Notion's customers used an AI feature built on top of Fireworks.

Two details stand out. First, the stolen transcripts include model outputs as well as user inputs, so prompts and generated answers are both in scope. Second, Notion's statement that much of the data was unattributable suggests the records were not cleanly tied to individual accounts, which complicates both notification and the assessment customers would want to make about whether their own content was involved.

The stated motive is also worth unpacking. Reselling access to LLMs generally means obtaining API keys or equivalent credentials and offering them to third parties who then run inference at someone else's expense. In practice that can look like stolen keys sold on underground markets, or proxy services that quietly route traffic through compromised accounts. Billing information being taken alongside datasets fits that pattern, since account and payment details can support fraudulent use or resale. This is my reading of the described motive, not a claim about what specifically was taken beyond the billing information and datasets Fireworks acknowledged.

The broader lesson concerns supply chains. Notion did the right things in sequence, disconnecting, deleting and severing connectivity once Fireworks raised the alarm, but the underlying exposure had already occurred at a layer Notion did not control. Product teams that build on third-party inference providers inherit that provider's security posture. Practical responses include minimizing what sensitive content is written into prompts and logs, scoping API credentials tightly and rotating them, and having a rehearsed plan for cutting a vendor off quickly.

Why it matters: Any team routing user content through a third-party model provider is exposed to incidents at that provider, not just at its own perimeter. The Notion case shows that AI feature logs can contain full conversational content, so prompt handling and vendor contracts deserve the same scrutiny as primary data stores. Expect more questions from security reviewers about which inference providers sit behind customer-facing AI features and what data they retain. That last point is my inference from the facts reported.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

This is Fireworks' second data breach in as many months.