Venkata Phani Patelkhana

AI agents may begin a task with a valid identity, approved tools, limited permissions, and clear instructions. But the situation can change while the task is running.

New user input, retrieved documents, tool responses, memory, or messages from other agents can influence what the agent decides to do next. An action that appeared safe at the beginning may become sensitive or risky later, even when the agent still has valid access.

Using a simple hypothetical story, this talk explains why initial security checks are only the starting point. It introduces three practical ideas: ongoing risk checks, connected visibility across the agent’s journey, and guardrails that respond when the level of risk changes.

The session follows the path from the user’s original request to the agent’s decision, tool use, and final action. It also shows how a system can respond by reducing access, requesting human approval, limiting the agent to read-only tasks, or stopping a sensitive action.

The main message is simple: trust should not be treated as permanent. When an agent’s context, behavior, or intended action changes, the risk should be checked again before the action is completed.

Biography:

Venkata Phani Patelkhana is an IEEE Senior Member and Distinguished Engineer at Dell Technologies, with more than 19 years of experience in enterprise API platforms, distributed systems, API security, and agentic AI architectures.

His current work focuses on AI control planes, MCP gateways, AI Security, agent identity, runtime authorization, and helping enterprises transition safely from API-first to agent-first ecosystems. He holds patents related to decentralized API gateway architectures and regularly writes and speaks about practical approaches to securing and governing enterprise API and AI ecosystems.