Hanselminutes with Scott Hanselman Trusting Agentic AI with Dr. Dawn Song
Dec 18, 2025
In this engaging conversation, Dr. Dawn Song, a computer science professor and MacArthur Fellow, dives into the world of agentic AI and its implications for security and privacy. She discusses her transition from physics to security, the impact of prestigious awards on her research freedom, and the mission of the Berkeley Center for Responsible Decentralized Intelligence. Dawn emphasizes the challenges of securing AI agents, the risks of complex systems, and her personal journey building custom agents for increased productivity.
AI Snips
Chapters
Transcript
Episode notes
Agentic AI's Rapid, Long-Built Rise
- Agentic AI has long development roots and progress accelerated faster than many expected.
- Researchers were surprised by rapid frontier advances after key model releases like GPT iterations.
Don't Confuse Transformers With The Only Future
- Transformers and next-token prediction proved powerful but are not the only path forward.
- The community should explore diverse architectures because current paradigms still have major limitations.
Agentic AI Lives On A Flexibility Spectrum
- Agentic AI spans a broad spectrum of autonomy, tools, and flexibility, not a single definition.
- As agent flexibility increases, the attack surface and security complexity also grow.