New turn in Fireworks AI breach after cloud provider hack

The topic is a series of breaches at Fireworks, a company that operates open source AI models for developers. In October, attackers took API keys and tokens for a small number of accounts.

Fireworks reported a second breach in an advisory. It tells that a cloud provider that kept its research machines had an incident on 23 October. On 24 October a party with no approval used it to get access to Fireworks development machines and operate a workload. Fireworks tells that it found the work the same day. The source gives no number of customers that it hit.

What the attackers took

Fireworks tells that the attackers got:

• resource metadata, such as names and descriptions
• billing data, such as addresses, invoice history and account balance
• parts of model weights and datasets for some customers

Fireworks tells that they could not get card numbers, because it does not keep them.

The proof

The source gives no test results and no count of customers. Fireworks tells that on 25 October it saw no more access on the platform its customers use. It replaced credentials and isolated the systems that the attack hit.

What is not known

Fireworks works with Unit 42, an incident response team at Palo Alto Networks, to find the full scope. The source does not say if the same group did the two attacks. In the first attack, Fireworks told the signs point to a group that sells model access that it took without approval.

Insight: The way in moved to the supplier

Wiz tells that 9 or more AI developer platforms had breaches from 6 to 30 October, and that attackers went for keys. Fireworks is one of the 9. Its second breach came through a cloud host of its research machines. This looks like attackers who move to a less safe supplier, where the same value waits. Your risk includes the suppliers of your vendors.