On August 19, 2026, Capria Ventures hosted an AMA session titled “Building Secure AI Products: The New Threat Model” featuring Akash Mahajan, Founder & CEO of Appsecco and Kloudle and a cybersecurity expert focused on securing AI applications, agents, and MCP systems.
The session explored how the shift from traditional software to agentic systems is creating a fundamentally different security landscape. Through practical demonstrations, Akash showed how AI agents can turn existing software vulnerabilities into significantly more powerful attack vectors when they are given access to tools, data, and the ability to take actions autonomously. The discussion covered topics including prompt injection, role-based access control, MCP security, authentication and authorization, sandboxing, deterministic controls, defense in depth, and the risks of allowing user-controlled data to enter an agent’s context.
Key takeaways included the importance of thinking about security from the very beginning, treating agents as trusted insiders, enforcing permissions at the tool and action level, separating untrusted data from executable instructions, and assuming that prompt injection will remain possible rather than relying on the model itself as a security boundary. The session reinforced that as AI agents become more capable, security must become a core part of product architecture: combining traditional security practices with strong authorization, sandboxing, threat modeling, and deterministic controls to ensure that even if an agent is manipulated, it cannot perform actions it was never authorized to take.
