Posted in Incident Response
Fireworks AI was breached twice in three weeks. The second time, the way in was someone else's cloud. From…
Craig M. — Chief Product Officer @ Cyberr®
Fireworks AI was breached twice in three weeks. The second time, the way in was someone else's cloud.
From the company's own update (9 Nov): in early October, an intruder used internal credentials to capture secrets from evaluation jobs. Fireworks rotated keys and tightened access.
Then on 23 October, a third-party cloud provider hosting Fireworks' R&D machines had its own incident. The next day, the attacker used that foothold to reach Fireworks development machines and run a workload that exposed billing data and, for some customers, datasets and incomplete model weights.
Fireworks links the actor to a financially motivated "LLMJacking" group: steal access, resell model tokens. That is a recurring revenue motive. Rotating a key removes the key. It does not remove the attacker's reason to come back.
My takeaway: your blast radius is set by your least-governed compute, and R&D is usually it.
What I'd check this week:
→ Inventory every place dev and research workloads run, including GPU and niche cloud providers outside your main tenant.
→ Map what those machines can reach. If a dev box can read customer datasets, it is production.
→ Treat a hosting provider's security notice as your own incident: rotate credentials and tokens that live on that infrastructure the same day.
→ After "containment", keep hunting for the same actor through different doors for at least 30 days.
Honest question for security leaders: who in your org owns the security of the GPU box someone rented last quarter?
#CloudSecurity #ThirdPartyRisk #AISecurity
4 reactions · 0 comments
Topics: #thirdpartyrisk, #cloudsecurity, #aisecurity