Criminals broke into Fireworks AI twice in October. The first time they took the API keys that some customers had stored there. The second time, through a cloud provider that hosted the company's research machines, they reached billing details and, for some customers, incomplete models and datasets. Fireworks has switched off self-serve dataset uploads and downloads for batch jobs and dataset-based training while it investigates. Its normal inference service keeps no prompts or outputs by default, and it has found no sign of impact there.
The datasets and billing records are the part that sounds serious. The keys are the part that costs money.
A key behaves like cash
An API key is a long string of characters that your app shows an AI company to prove it may use your account. Nothing else is checked: no second code, no phone, no name. Whoever holds the string spends on your bill.
The contracts are written the same way. The standard terms of the big model providers make the account holder responsible for all activity under the account, and a refund is up to the provider. After the October thefts, some providers applied their fraud refund rules and at least one kept a no-refund policy, so a stolen key can mean a loss that you carry yourself.
One platform holds thousands of keys
A person who leaks a key on GitHub gives a thief one key. A platform that stores keys for its customers gives a thief all of them at once, and in October nine AI developer platforms reported break-ins of this kind. At one of them, Railway, the attacker's database searches looked for the names that AI keys usually have. Railway estimates the stolen keys led to $740,000 in unauthorized charges.
The speed comes from what happens next. The 2013 thief had to rent machines and wait for coins to trickle in, and the bill grew slowly. Today a thief passes the key to a gateway that pools stolen keys and resells access, and the buyers are already waiting. One service advertised unlimited access to top models for about $60 a month. In one case an actor sent more than 500,000 requests through such a gateway in under eight hours, which left the key's owner with nearly $1 million in charges, about $2 for every request.
What a small builder can do about it
Picture a shop owner who built a tool that drafts replies to customer emails. It costs about $15 a month, and its key is stored on a hosting platform along with the rest of the tool's settings. If that platform is broken into, the key goes to a gateway and is used all night. With no cap, the owner finds out from an email from the provider the next morning, and the amount is whatever the gateway's customers managed to use. With a $50 monthly cap at the provider, the worst night costs $50 and a new key.
That example comes down to a few habits that take an hour to set up:
Set a spending cap on every AI account you pay for. Give each tool its own key, so you can cancel one without breaking the others. Keep keys only where a tool needs them to run. And when any platform you use emails about a break-in, replace its keys that day, including keys you think nobody used.
The next breach of an AI platform will be announced as a data breach, and the first paragraph will be about datasets. For anyone building with AI, the better first question is which keys were stored there and how much each one was allowed to spend.