Jason Clinton, deputy CISO, Anthropic. Interviewed by Shannon Yavorsky. Clinton leads security at Anthropic and focuses on AI safety, cybersecurity and infrastructure. 

Key takeaways

  • AI capability growth remains structurally predictable 
    • Compute used in training models has grown exponentially for decades, with ~4x year-on-year increases expected to continue in the near term. 
    • This underpins high confidence that models will become materially more capable over the next two years. 
  • Implication: software economics may collapse—but value persists elsewhere 
    • Clinton suggests the marginal cost of software could approach zero, driven by AI-generated code and automation. 
    • However, value will shift to branding, convenience and especially data moats, which remain hard to replicate (e.g. security platforms with large deployed datasets). 
  • AI companies are evolving into ‘neo-hyperscalers’ 
    • Frontier model providers are positioning themselves as infrastructure platforms, offering “intelligence as a service”, analogous to cloud computing. 
    • The long-term opportunity lies less in consumer features and more in enterprise APIs and backend integration. 
  • Cybersecurity is entering an AI-driven arms race 
    • AI models are now capable of offensive techniques previously limited to nation-state actors, lowering the barrier for attackers. 
    • Defensive strategies must therefore layer multiple controls (“Swiss cheese model”) augmented with AI at each layer. 
  • Agents are already transforming enterprise workflows 
    • Internal agents can autonomously diagnose outages, generate fixes and coordinate teams via existing tools (e.g. Slack), reducing resolution times to minutes. 
    • Use cases extend across legal, product, partnerships and data analysis, often replacing traditional data science workflows. 
  • Human oversight remains central—but is shifting 
    • Systems increasingly operate with “humans on the loop” rather than “in the loop”, with humans acting as final approvers rather than active operators. 
    • Legal frameworks (e.g. copyright) still hinge on human involvement, leaving open questions as autonomy increases. 
  • Mission alignment is a competitive advantage 
    • Strong organisational alignment around principles (e.g. restrictions on harmful use cases) drives better decision-making and execution across teams. 

Takeaway for CVCs

The investable edge is shifting away from standalone software and towards infrastructure, data moats and security capabilities in an AI-accelerated environment where both opportunity and risk are compounding rapidly. 

This is an AI-generated summary, which has been lightly edited by GCV staff.