Mr. Prabhu Sivapunniyam

Title of the Talk:
When AI Becomes the Attack Surface: Securing Enterprise AI Agents

Abstract:
Enterprise AI is evolving from systems that generate answers to agents that take action. By retrieving information, invoking tools and coordinating workflows, AI agents are expanding what organizations can automate. But this growing autonomy introduces a fundamental security question: what happens when an AI system can act on our behalf—and how do we limit the damage when it gets something wrong?
Traditional security controls remain essential, but agentic systems introduce additional trust boundaries. Agents interpret natural-language instructions, consume external content and use retrieved context to guide their decisions. A malicious instruction embedded in a document, an integration with excessive permissions or an unsafe action passed between agents can expose sensitive information or trigger unintended operations. The challenge extends beyond protecting the model to securing the data, tools, memory and execution paths surrounding it.
This talk explores these challenges through the lens of enterprise architecture, AI-powered security validation and multi-agent orchestration. It examines how prompt injection, untrusted retrieval content and excessive agency can influence agent behavior, and how scenario-based testing can help teams evaluate whether safeguards hold under adversarial conditions.
The session presents a practical approach to securing enterprise agents through narrowly scoped permissions, explicit trust boundaries, validation before execution, human approval for sensitive actions and end-to-end tracing. It also explores how simulated security scenarios and detection validation can help teams assess both prevention controls and their ability to recognize failures.

Bio:
Prabhu Sivapunniyam is a Principal Member of Technical Staff at Salesforce, specializing in artificial intelligence, cloud architecture, distributed systems and enterprise data platforms. He brings more than 25 years of software engineering experience, with current work spanning AI-powered security validation, Retrieval-Augmented Generation (RAG), multi-agent orchestration and intelligent security automation. At Salesforce, he develops platforms that combine large language models, security scenario generation, real-time event simulation and detection pipeline orchestration. His work also includes context-aware agent frameworks, knowledge retrieval and human-guided automation for security workflows.
Previously, Prabhu worked at Microsoft, developing large-scale Azure Commerce services and cloud-based supply chain analytics platforms. His experience spans scalable data processing, production reliability, monitoring and governance. He holds a Master of Computer Applications and has led technical delivery and mentored engineering teams across enterprise environments.

.