Fireworks AI hacked in coordinated attack on inference providers; partial model weights stolen

Inference providers under coordinated attack; Fireworks partially compromised

A coordinated wave of hacks is targeting inference infrastructure. Fireworks AI, Railway, and OpenCode have all faced breaches in recent weeks, suggesting attackers are systematically moving through startup-scale compute providers.

The Fireworks incident is notable because attackers only partially exfiltrated model weights before being caught the same day. They were unable to steal a complete copy. Fireworks hosts open models that have been post-trained by various companies, making the weights a valuable target for theft.

The pattern suggests roving attacks. Once one inference provider hardens defenses, attackers move to the next target in the category rather than persisting against a single hardened system.

Precedent and systemic risk

Ilya Sutskever flagged this risk in September, predicting that neo-clouds with limited cybersecurity could become targets for rogue agents attempting to spawn additional compute instances. While this Fireworks incident doesn't involve agent takeover—the attackers were human-directed—it validates the underlying vulnerability: inference providers are an attractive attack surface for both data theft and potential compute hijacking.

Sutskever argued that infrastructure companies with strong security practices should help neo-clouds strengthen defenses. The Fireworks breach suggests that argument has already become operational concern rather than theoretical threat.