When Nobody's in Control: AI, Cyber Risk, and the Economics of Digital Trust
Summary
Daleep Singh discusses the dark side of AI with Shuman Ghosemajumdar, a former Google security executive, exploring how AI lowers the barrier for cybercrime through automated deepfakes and "smart cow" vulnerabilities. They analyze the Mythos incident and the futility of centralized AI controls, arguing that on-device verification is essential to restore trust in digital communications. The conversation also covers the AI capex cycle, suggesting that diminishing returns on model size will limit the productivity boom and lead to a centralized tech oligopoly.
Topics discussed
Introduction and the dual nature of AI
Guest background: Schuman Gossam from Google
Early Google: Secrecy and automation origins
Evolution of cyber threats from manual to AI
AI-enhanced fraud and the economics of crime
The trust deficit in digital communication
Solution: On-device verification and privacy
Network effects and the 'Blue Bubble' analogy
State-sponsored attacks on water infrastructure
The 'Smart Cow' problem in cybersecurity
AI lowering barriers for amateur attackers
Mythos incident and export control lessons
Jailbreaks and the limits of model alignment
Human error and the 'PEBCAK' factor
Geopolitical leverage and open-source security
AI valuations and the transformer asymptote
Productivity booms and labor market displacement
Competitive dynamics and ecosystem advantages
Political backlash and autonomous systems
The paperclip maximizer and goal misalignment
Optimism and societal dialogue on AI
Listen ad-free on Castria