Security as Architecture: Building Agents That Can Be Stopped | Piyush Gupta

Piyush Gupta has been building in AI since 2005, and now leads product for the Enterprise AI Platform at Uniphore, where he works on the models and agentic layers that Fortune 100 companies are deploying. He has been a CPO three times with two exits, and before that founded a mapping company in India that Google tried to acquire, and built a mobile banking product that brought millions of people into the financial system for the first time.

In this episode, we cover:

• Why he thinks about security as an architecture rather than a guardrail, and what a launch blocking spec for an agent actually looks like
• The kill switch question most agentic platforms have not answered: how fast can you pause an autonomous workflow
• Why the model was not the problem in the Hugging Face incident, and why Sam Altman's response is the more worrying part
• The difference between technology and product, and why we are deep in a technology cycle right now
• Building a way to send money by missed call in India, taking a year to reach 100 users, then $10M in payments the year after
• What the World Bank found about EKO's customers that brought him to tears
• Why open weight model adoption is about to accelerate, and how model training becomes part of the product development lifecycle
• Why evals at scale in production is still the unsolved problem blocking enterprise agents

A conversation for PMs and product leaders building AI products who want to understand what actually breaks between proof of concept and production, and how to design for it before you ship.