Harshit Kohli on AWS, MCP, AI Agents, and Secure Enterprise AI

What this episode covers

Harshit Kohli is helping enterprises understand where AI, cloud, streaming analytics, and cybersecurity are heading next. As a Senior Technical Account Manager at Amazon Web Services and a GenAI streaming specialist, Harshit works with customers on cloud adoption, artificial intelligence adoption, and real-time data use cases. He is also pursuing a doctorate in artificial intelligence and continues to serve as one of BIG’s All-Star Judges.In this episode, Russ and Harshit discuss what he is seeing across AI, cybersecurity, MCP, and agentic systems after speaking at industry events including RBLN and the MCP Dev Summit. Harshit explains why companies are moving quickly toward AI, but must also keep security, governance, authorization, and authentication at the center of every deployment.The conversation dives into model context protocol, or MCP, and how it is evolving beyond simple request and response workflows. Harshit shares how he demonstrated a streaming MCP architecture using Amazon MSK, WebSockets, IAM authorization, and Amazon Bedrock to push real-time context into AI agents. The goal is to help organizations detect anomalies, understand root causes, and act faster when minutes or seconds can carry major business risk.Russ and Harshit also explore the risks around compromised MCP servers, agent hijacking, shadow AI, least privilege access, sandbox execution, human in the loop validation, and the growing need for governance around AI agents that can now plan, act, execute code, and manage infrastructure.Along the way, Harshit shares why enterprises should not chase AI for its own sake, why many projects may fail without clear use cases and controls, and why the best companies are not replacing people with AI, but investing in their teams so people can do more with it.Topics Covered:[00:00] Welcome and intro, Harshit Kohli and BIG’s All-Star Judges[00:43] Harshit’s role at AWS and work in GenAI streaming[01:55] His doctorate in artificial intelligence and industry speaking[02:19] Key themes from RBLN around AI, cybersecurity, and governance[03:56] AI-driven phishing volume and security industry response[04:33] Why governance and security must be built into AI products[05:44] MCP Dev Summit and the evolution of model context protocol[06:23] Moving MCP beyond basic request and response[07:00] Harshit’s streaming MCP demo with Amazon MSK, WebSockets, IAM, and Bedrock[08:26] Why real-time anomaly detection matters in financial systems[08:53] Root cause analysis and remediation through AI agents[10:29] Observability, telemetry, and cost anomaly detection[12:16] MCP design flaws, exposed servers, and enterprise risk[12:52] Why AI agents now execute code and manage workflows[14:41] What a compromised MCP server means[15:05] Sandbox execution and least privilege access[17:23] Consent, verification, and human in the loop controls[18:40] The rise of enterprise AI agents[19:21] Why agent autonomy creates governance challenges[20:38] Deploying agents versus deploying applications[22:38] Agent hijacking and prompt injection risks[23:00] Poisoned documents, malicious tools, and compromised MCP servers[25:19] Patch windows, enterprise readiness, and AI cyber risk[26:30] Shadow AI inside organizations[26:55] Employees using unapproved AI tools at work[28:57] Data leakage, compliance violations, and hallucination risk[30:46] Why some AI projects may be canceled[31:07] Starting with real use cases instead of forcing AI[32:44] Whether AI has become infrastructure yet[33:39] What the best companies are quietly getting right[34:46] Final thoughts on investing in people while adopting AI

PodParley-generated summary based on available episode metadata and transcript content.