Fireworks AI breached twice in a month as AI-powered hacking overwhelms security firms
A second breach at AI hosting firm Fireworks exposed billing and model data, as Mandiant, CrowdStrike and Kroll say AI-driven attacks are swamping their response teams.
What happened: Fireworks AI, a platform that runs open-source AI models like Llama and Mistral for developers, disclosed its second security breach in about three weeks. In early October, attackers stole API keys from some accounts. After Fireworks revoked those credentials, it found a separate intrusion on October 23: a hacked third-party cloud provider let attackers reach Fireworks' development machines and run an unauthorized workload touching more customer data. Separately, NBC News reports that top incident-response firms Mandiant, CrowdStrike and Kroll are so swamped with AI-accelerated attack cases that they are turning away some new victims seeking emergency help.
Why it matters: The breach exposed account IDs, billing and shipping addresses, invoice and payment history, credit usage, and incomplete model weights and datasets for some customers — though not payment card numbers, which Fireworks doesn't store. That data can fuel further fraud or targeted follow-up attacks. More broadly, if the firms companies call during a cyberattack are themselves stretched thin because attackers now use AI to multiply how many targets they can hit at once, victims may wait longer for help exactly when speed matters most.
How it works, plainly: Investigators linked the October incident to a financially motivated group running 'LLMjacking' — stealing API keys and tokens, then reselling unauthorized access to AI models through illicit routing gateways. The second breach shows how one weak link cascades: a cloud vendor's own breach gave attackers a foothold into Fireworks' internal dev machines, which then exposed customer account data. Google's threat intelligence team recently found LLMjacking has already driven fraudulent cloud-AI charges as high as $1 million at other victim organizations, since stolen keys let criminals run AI workloads on someone else's bill undetected for a while.
The rollout: Fireworks says it detected the October 24 intrusion the same day, revoked and replaced credentials, isolated affected systems, and tightened access controls; as of October 25 it had found no further unauthorized access. It has also brought in Palo Alto Networks' Unit 42 to investigate the full scope. Meanwhile, Mandiant, CrowdStrike and Kroll executives are publicly flagging a capacity crunch: client demand for emergency response is rising faster than their ability to staff it, a strain they attribute directly to attackers automating intrusions with AI.