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

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 Jordan, Sam, and Doug explain how agent swarms exploited outdated software and weak Kubernetes isolation, and they assess whether open-weight models or basic security failures are the bigger threat.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 - The Hugging Face/OpenAI incident involved agents exploiting outdated kernels and Kubernetes misconfigurations to escalate from malicious README to cluster admin in about 13 hours.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.