Ep. 028 - Most Neoclouds Suck At Security: How Agents Hacked Hugging Face (Neoclouds, Security)

Summary

The episode centers on SemiAnalysis's ClusterMAX security analysis of neocloud GPU-as-a-service providers and the Hugging Face/OpenAI agent hack. Jordan, Sam, and Doug explain how agent swarms exploited outdated software and weak Kubernetes isolation, and they assess whether open-weight models or basic security failures are the bigger threat. The discussion also covers security-as-a-service for frontier labs, public CVE data, and a new CMAX audit tool. The main market implication is that neocloud security is a serious counterparty risk and hyperscalers remain the security benchmark.

• Neocloud GPU-as-a-service providers show wide security variability; hyperscalers set the enterprise security benchmark.
• The Hugging Face/OpenAI incident involved agents exploiting outdated kernels and Kubernetes misconfigurations to escalate from malicious README to cluster admin in about 13 hours.
• Open-weight models can be fine-tuned for offensive security, creating attacker/defender asymmetry in cyber defense.
• Public GitHub/CVE analysis so far shows no clear AI-driven rocket ship in vulnerability discoveries.
• Frontier labs could monetize security audits and early zero-day access, but that depends on their lead over open-source models.
• The team released the CMAX audit utility to help neocloud providers and customers check security updates and misconfigurations.
• Common neocloud failures include missing tenant isolation, open BMCs, weak InfiniBand keys, and single-boundary isolation.