Fireworks AI told customers on November 9 that an October 24 intrusion exposed billing records. For some accounts, datasets and incomplete model weights were also exposed. Crypto Briefing reports the company first detected unauthorized access on October 6. This marks Fireworks' second security incident in recent weeks.
What the attacker reached
RuntimeWire reports the October 24 breach came through a cloud provider and hit development machines. The exposed data includes resource metadata, billing and shipping addresses, invoices, payment histories and account balances. Fireworks says it does not store payment card numbers. For a limited set of accounts, the attacker also reached stored secrets and datasets. Incomplete model weights were also exposed.
Bloomberg reports the breach included tens of thousands of files. Startup Fortune reports Fireworks attributed the breach to LLMjacking. The company is working with Palo Alto Networks' Unit 42. Fireworks serves about 10,000 customers and was valued at $17.5 billion in July. Only a couple of clients have left, because replacement GPU capacity is hard to find.
The resale market
Google Threat Intelligence Group reported on November 2 that threat actors are stealing cloud credentials and API keys to hijack access to AI models. They then resell that access. One actor made over 500,000 LLM API calls in less than eight hours. Help Net Security reports some victims faced bills up to one million dollars. The attackers eventually lost access.
The Hacker News reports UNC6993 breached a cloud instance through a misconfigured service account key. The group quickly racked up over $100,000 in billing charges. SiliconANGLE reports another actor scraped LLM API keys from public GitHub repositories. Each key was tested with a request capped at a single output token.
The billing ledger you keep
Atelier calls nine provider APIs directly. Its usage ledger records the cost and time of each request. The ledger is how I would spot LLMjacking if someone grabbed a key. I would check whether the ledger matches the bill. A gap between them means the tokens went somewhere else.
AI Trader stores every decision with its inputs, output and cost. I would reconcile broker statements against those records. Keep Fix Cut tracks anonymous session usage and token cost for 30 days. If the bill spiked, I would check the session count first. Then I would check the cost per scan.
The ledger is what you spent.
Pasquale Pillitteri reports Wiz counted at least nine AI developer platforms that disclosed an incident in October. Three of them saw signs of an agent swarm. Fireworks says it has not seen any impact on its Inference service, Training SDK or Nexus. Those services run with zero data retention by default. OpenCode's breach exposed API keys in a legacy system being phased out. That cleanup took longer than the keys stayed safe.