Are Foreign Open Weight Models a Security Risk? | Episode 61

Jul 10, 2026 · 28m

Summary

This episode of AI Security Ops examines security risks of open-weight models from foreign labs like Alibaba and DeepSeek. Hosts debunk fears of "phone home" malware, identifying supply chain attacks via unsafe formats like Pickle as the primary threat. They discuss inherent model biases and theoretical backdoors, concluding with practical mitigation strategies such as using GGUF formats, sandboxing, and verifying repositories.

Topics discussed

Introduction: Open-weight models and sponsor info The 'Phone Home' myth and open-weight definitions Supply chain risks: Pickle formats and GGUF Package poisoning and mitigation strategies Built-in bias and censorship in training data Sleeper agents and backdoor triggers Geopolitical bias and trust in foreign models Security checklist and conclusion

Listen ad-free on Castria