Fireworks AI widened its breach: billing data, datasets and model weights in attackers' hands
Direct answer
On November 9, 2026, Fireworks AI, an inference and fine-tuning platform with more than 10,000 customers — including Notion and Cursor — widened its October breach report: attackers reached billing data from its accounts and, in a limited set of accounts, stored secrets, datasets and incomplete model weights. The company says it does not store payment card numbers and that those were not affected. Inference, Training SDK and Nexus services were not affected, according to the company, because they retain no data by default. The update changes the story: the first bulletin on October 9 said there was no evidence of access to models, datasets or training data.
How the breach unfolded
- October 6: Fireworks detected unauthorized use of internal credentials. The first bulletin, on October 9, spoke of stored secrets in evaluation jobs and no evidence of access to models or datasets.
- October 23 and 24: an outside cloud provider hosting Fireworks research and development machines suffered its own security incident. The next day, the intruder used that opening to reach the development machines and run an unauthorized workload. Fireworks says it detected that activity the same day and has seen no further successful access on its customer platform after October 25.
- November 9: the update adds billing data (billing and shipping addresses, invoice and payment history, balances and credit usage) and, for a limited set of accounts, secrets, datasets and incomplete model weights. The company did not publish how many accounts were affected or the volume of data.
- The motive: Fireworks links the activity to a financially motivated group that sells stolen access to language models through illicit routing gateways, a scheme known as LLMJacking. The company cites November 2 research by Google Threat Intelligence Group and Mandiant. The company engaged Palo Alto Networks Unit 42 and notified law enforcement.
What was affected and what stayed out
- Billing data: addresses, invoice and payment history, balances and credit usage for the affected accounts. Fireworks says it does not store card numbers, so those were not in its systems.
- Incomplete model weights and datasets: only in a limited set of accounts. An incomplete weight is still the customer's intellectual property: a checkpoint or an unfinished training artifact, with value even though it is unfinished. A fine-tuning dataset can hold customer records or internal files.
- Services with no impact according to the company: Inference, Training SDK and Nexus, which retain no data by default. The company says self-serve dataset uploads and downloads for Batch Inference and dataset-based training stay disabled while the investigation continues.
- The cost to customers: Bloomberg reported, citing a person familiar with the matter, that the intrusion included tens of thousands of files. Forbes reported that two customers already moved substantial workloads off Fireworks, while two others said they were unaware of any hack.
What to demand from your AI vendor after this case
- Ask for zero retention by default: the Fireworks services that store no data were not affected. When you evaluate a vendor, ask what data it stores, where and for how long. Zero retention is a security feature: ask for it in writing.
- Demand breach notice in writing: Fireworks notified affected accounts directly, but the real scope became known five weeks after the first bulletin. Your contract should set notice deadlines and scope updates, not just a first email.
- Separate your credentials by environment: the breach came through internal credentials and an outside provider. Development and production credentials must live apart, with rotation and least access.
- Have an exit plan: two customers moved their workloads within days. If your operation depends on a single inference vendor, an incident like this leaves you with no plan B. Document how you would migrate your models and data before you need to.
Note: this article summarizes Fireworks AI's security bulletins of October 10 and November 9, 2026 and coverage in Silicon Report, AI Secure Platform, Breach History and Vanda Team. Customer figures are the company's published figures. The number of affected accounts and the volume of data have not been published. Fireworks' conclusions are provisional according to the company itself.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- AegisAI Proteus claims to detect 99% of phishing: the AI vs AI war in your inbox
- Privacy and data: what happens to your customers' information
- OpenAI expanded its agent monitoring: in a replay of the incident, the alert would have come 30 hours before the breach
Frequently asked questions
What data was affected in the Fireworks AI breach?
Billing data (addresses, invoice and payment history, balances and credit usage) and, in a limited set of accounts, stored secrets, datasets and incomplete model weights. Card numbers were not affected because Fireworks says it does not store them.
Why does the November 9 update matter so much?
Because it changes the first report: the October 9 bulletin said there was no evidence of access to models, datasets or training data. Five weeks later, the company confirmed there was access to datasets and incomplete weights in some accounts.
What is LLMJacking?
It is a scheme in which financially motivated actors steal access to language models and resell it through illicit routing gateways. Fireworks links its breach to this scheme, citing research by Google Threat Intelligence Group and Mandiant.
Are my models safe if I use an inference vendor?
It depends on what the vendor retains. In this case, the services with zero retention by default (Inference, Training SDK and Nexus) were not affected. Ask your vendor what data it stores and demand breach notice in writing in the contract.
