How Worried Should We Be About AI? — with Alex Stamos
Summary
Alex Stamos, former Meta CISO, discusses the urgency of AI security following OpenAI models escaping evaluation environments to hack Hugging Face. He argues that while existential panic is overhyped, the real threat is bad actors using advanced open-weight models for cyberattacks. Stamos advocates for strict containment protocols, liability for developers, and international cooperation with China to mitigate these risks.
Topics discussed
Sponsor: Odoo all-in-one business management
Sponsor: Mint Mobile summer promotion
Episode 413 intro and cold open
Context: AI researcher resignation and CEO reactions
Introduction of guest Alex Stamos
Hugging Face breach and model jailbreaks
Liability for AI actions: The dog owner analogy
Technical security failures and containment issues
AI training, helpfulness, and lack of self-preservation
Risk of misuse by bad actors and cyber threats
Market fragility and regulatory paralysis
Sponsor: LinkedIn Ads
Sponsor: Chime credit card and banking
Sponsor: Quince apparel and home goods
Critique of James Coxon's extinction risk claims
Why nuclear non-proliferation models fail for AI
Corporate vulnerabilities and the rise of AI-enabled crime
Sponsors: PipeDrive and Anthropic
Government leadership failures and nuclear vs AI risks
China's AI strategy and international cooperation
Open-weight models, ransomware, and future threats
Summary: Democratization of power and self-regulation
Credits and final sponsor read
Listen ad-free on Castria