Fireworks AI says breach reached billing data, customer datasets and incomplete model weights

What Fireworks AI disclosed

Fireworks AI, which runs a cloud platform for serving and customizing open and custom AI models, updated its security bulletin on 9 November 2026 and told customers that an incident first disclosed in October was broader than it had said. The company now says attackers reached billing and account records, including billing addresses, payment histories and credit usage. For a limited set of accounts they also reached uploaded datasets and what the company calls incomplete model weights. Fireworks says it does not store payment card numbers and that card data was not affected.

The bulletin was not reachable from our systems, so the details of the update come from the company's notice as quoted by several outlets, among them Cybersecurity News and The Defiant, and from reports by Bloomberg and The Information. Fireworks has not said how many customers are affected, beyond calling the group limited. Bloomberg reported that the stolen material included tens of thousands of files, some containing customer data; the company has not confirmed that figure.

How the incident unfolded

Fireworks first detected trouble in early October, when someone used internal credentials without authorization. According to the original bulletin, the attacker reached environment variables in evaluation jobs for a limited number of customer accounts, which can hold API keys and other secrets. At the time the company said it had seen no evidence that customer models or datasets were accessed, began containment the same day and notified the customers it identified.

The 9 November update describes a second path. According to the company's notice as reported, intruders entered on 24 October through a third-party cloud provider that hosts Fireworks research and development machines, and from there reached additional customer information. Fireworks has not named the provider. It says it revoked affected keys and credentials, notified law enforcement and is working with Palo Alto Networks' Unit 42 incident response team. It has also switched off self-service dataset uploads and downloads used by batch inference and dataset-based training while the investigation continues. The company says it has detected no impact on its Inference service, its Training SDK or Nexus, which do not retain customer data by default.

Who is affected and what customers should do

The Information reported that Fireworks told large customers including the legal AI company Harvey that their data had been exposed, and that some customers are considering moving their workloads elsewhere. Neither Harvey nor Fireworks has commented publicly on the specific accounts. Fireworks advises affected customers to replace any keys it revoked, rotate third-party credentials that may have sat in their environment variables or jobs, and review which datasets they uploaded for batch jobs and fine-tuning. Customers that did not receive a notice were not identified as affected, the company says, and it has set up a dedicated security contact for questions.

Partial model weights matter because fine-tuned models can encode a customer's private training data and the work that went into a custom model. Datasets uploaded for fine-tuning often contain the same proprietary documents, conversations or code that companies keep off public services. The exposed billing data adds a phishing risk, since attackers can write convincing invoices or payment notices to known customers.

Why it matters for AI infrastructure

Fireworks raised $1.505 billion at a $17.5 billion valuation in July and is one of the largest independent inference providers. It was already one of at least nine AI developer platforms that Wiz counted as breached in October, in a report on stolen keys and LLMjacking, where attackers harvest provider credentials to resell model access. The new disclosure shows the risk reaching beyond keys to the training data and model artifacts that customers hand to inference platforms.

Open questions are how many customers lost datasets or weights, which cloud provider was the entry point, whether the attackers have been identified and whether regulators will receive breach notifications in the EU or US states. More security coverage is in the cybersecurity hub.

Questions people ask

What data was exposed in the Fireworks AI breach?

Billing and account records such as billing addresses, payment histories and credit usage, and for a limited set of accounts datasets and incomplete model weights. Fireworks says payment card numbers were not affected.

Was the Fireworks inference service affected?

Fireworks says it detected no impact on Inference, the Training SDK or Nexus, which keep no customer data by default.

What should Fireworks AI customers do?

Replace revoked keys, rotate third-party credentials stored in jobs or environment variables, and review datasets uploaded for batch jobs, following the company's notice.

Sources

1. October 2026 security incident (Fireworks AI, 9 Nov 2026)
2. Hackers Breach Fireworks AI Through Third-Party Cloud Provider to Steal Customer Data (Cybersecurity News, 9 Nov 2026)
3. Fireworks Says Breach Exposed Billing Data, Datasets and Incomplete Model Weights (The Defiant, 9 Nov 2026)
4. New Fireworks Hack Threatens $1 Billion Business (The Information, 9 Nov 2026)