The Lawfare Podcast Cybersecurity and AI
36 snips
Apr 3, 2023 
A lively panel explores the race to secure AI systems, the value of model weights, and fears of intellectual-property theft. They dig into how large language models could aid hackers, supercharge phishing and political disinformation, and help defenders spot vulnerabilities. The conversation also weighs AI’s limits, the risks of releasing powerful tools, and the need for auditing and red-team tests.
AI Snips
Chapters
Transcript
Episode notes
GRU Disinformation Campaign
- The GRU's Syria campaign involved fake journalists publishing in lefty magazines.
- Now, one person could manage such an operation using AI.
Stealing Linear Algebra
- LLMs like ChatGPT are a combination of code, model weights (values of connections), and training data.
- Protecting these models is crucial, similar to protecting other valuable software, but the model weights hold significant value.
Model Weights as Targets
- Model weights, 200GB-1TB files, are a critical target for adversaries like the PLA.
- Internal controls and data protection are crucial due to potential insider threats and espionage.