Fireworks AI breached via cloud provider: Notion data exposed
Second breach at Fireworks AI: through a third-party cloud provider, attackers took billing data, datasets and Notion AI chat transcripts.

Fireworks AI, a US inference and fine-tuning platform for open-weight models such as Llama, Mistral and Gemma, published an update to its security advisory on Monday, 9 November. On top of the incident discovered on 6 October, it suffered a second one at the end of October, starting from a third-party cloud provider that hosted its research and development machines. The second incident considerably widens the data involved: resource metadata, billing data and, for some customers, parts of model weights and datasets. Among the affected customers is Notion, which confirmed that Notion AI chat transcripts were exposed in fewer than 1% of workspaces.
The two incidents
First incident (October). On 6 October Fireworks detected unauthorised use of internal credentials and disclosed it on 9 October. The attackers had read the environment variables of evaluation jobs, where API keys and tokens belonging to a limited number of customer accounts were stored as secrets. Fireworks linked the attack to a financially motivated criminal group that resells access to language models through illicit gateways, known as LLMjacking. At the time, The Register reports, the company said it had no evidence of access to models, datasets, training data, inference traffic or personal data. The response: revocation of the compromised internal credentials, of the affected customers’ API keys and of the keys created by the intruders.
Second incident (23–24 October). On 23 October the cloud provider hosting Fireworks’ research and development machines suffered an incident of its own. On 24 October the attacker used that access to reach Fireworks’ development machines and run an unauthorised workload, which affected further customer data. Fireworks says it detected the activity the same day, closed the known critical vulnerabilities and, since 25 October, has seen no further successful access to the platform customers use. The name of the cloud provider has not been made public.
Since then the company has revoked and replaced credentials and service accounts, closed unauthorised network paths, isolated the affected systems and tightened access to data storage and analytics systems. The investigation is being handled by Palo Alto Networks’ Unit 42.
What data was exposed
According to the advisory, consistently reported by the outlets, the data falls into three groups:
- resource metadata: identifiers, display names, descriptions, creation times;
- billing: billing and shipping addresses, invoice and payment history, account balance and credit usage. Card numbers are not involved because Fireworks does not store them;
- model weights and datasets: for some customers, a subset of incomplete weights and datasets.
According to The Register, CEO Lin Qiao said the attackers do not appear to be after the weights and that most downloads were partial and stopped. Fireworks has not said how many customers are affected. Unlike the other sources, The Record already attributes the theft of account and billing data, in addition to secrets, to the first incident. According to Bloomberg, cited by The Register, Fireworks is assessing the impact with customers such as Cursor and Harvey: this does not mean their data was stolen.
The Notion case
Notion used Fireworks to serve the open-weight models in Notion AI and for the classification jobs of its Insights feature, which shows administrators how teams use AI. The timeline that emerges from the sources:
- on 2 November Fireworks warns Notion of an exposed dataset which, according to BleepingComputer, contained only synthetic test data;
- by 3 November Notion cuts traffic to Fireworks, deletes its data from the vendor’s systems and stops training runs;
- on 6 November Fireworks says malicious actors may have accessed customer data (The Record), and on 8 November it hands Notion the affected dataset (BleepingComputer).
The dataset contained Notion AI chat transcripts used for Insights: input and output written in the chat, without tool call results. For the vast majority of workspaces the chats are “unattributable”, meaning they carry no author or workspace, but their content may still identify the person writing. Notion has disabled Insights, no longer uses Fireworks for open-weight models and notes that the vendor had contractually committed to not retaining Notion data (zero data retention).
Techniques and attack chain
The sources do not list MITRE ATT&CK techniques. Based on the stated facts, the relevant ones are T1199 (trusted relationship: access came through the cloud provider), T1078.004 (valid cloud accounts, the internal credentials of the first incident), T1552 (unsecured credentials, the secrets in environment variables) and T1530 (data from cloud storage). Fireworks links both incidents to the LLMjacking wave described in early November by Google Threat Intelligence Group and Mandiant.
Impact for European and Italian companies
The issue is not only Fireworks but the supplier chain: a company may never have signed with Fireworks and still be involved, like Notion’s customers, because a SaaS service uses an inference provider that in turn uses a cloud provider. Chats and billing data may contain personal data: data controllers must assess, using the information received from the vendor, whether this is a breach to be notified to the supervisory authority within 72 hours (Art. 33 GDPR). For NIS2 entities, supply chain security is among the mandatory measures: AI providers and their sub-suppliers belong in the risk assessment like any other supplier.
What to do now
- Fireworks customers: check the notices you received, rotate API keys and tokens, especially those passed as environment variables to evaluation or fine-tuning jobs, and review usage and billing.
- Notion customers: if the workspace owner received the notice, use “Security incident review” to examine the affected chats and treat passwords, API keys and credentials pasted into Notion AI as exposed, so rotate them. Without a notice, according to Notion, no action is needed.
- Map your AI providers, direct and indirect: ask SaaS vendors for the list of sub-processors that handle prompts, chats and documents.
- No secrets in prompts or job variables: use a secret manager, short-lived tokens and scoped keys, with spending limits and usage alerts.
- Assess the exposure of datasets and models uploaded to inference or fine-tuning providers: intellectual property and any personal data they contain.
- Review your contracts: data retention, incident notification times, obligations towards sub-suppliers.
How ARXDOME helps
ARX Expose flags company accounts involved in breaches and services exposed on the Internet, useful when company credentials end up in a third-party incident. On monitored servers, ARX Sentry groups credential use and account change events, while ARX Horizon lets you query threat intelligence sources when a vendor publishes indicators of compromise.
Last updated:


