2026 GET-AI SERIES: 2 . Trust in AI Systems: Detecting, Defending, and Securing Intelligent Agents
We are excited to continue the Orange County Computer Society (OCCS) Global Emerging Technologies and Artificial Intelligence (GET-AI) Series—a monthly platform focused on transformative innovations in computer science and technology. Hosted by the IEEE Orange County Computer Society Chapter, this series brings together professionals, students, and tech enthusiasts to explore the cutting edge of what’s possible.
Following a highly engaging April session on Generative AI, where we explored LLMs, RAG, Agents, MCP, and hands-on AI application development, we are excited to bring you our May Tech Talk on “Security in AI.”
🔒 May Focus: Securing Generative AI
As AI systems evolve—from traditional models to LLM-powered agents interacting with enterprise systems and real-world tools—they introduce powerful capabilities along with new security challenges, including:
- Data leakage and prompt injection
- Model misuse and unauthorized access
- Risks in agent-driven automation
- Governance and compliance concerns
This session combines technical insights and practical demonstrations to explore how to build secure, trustworthy AI systems at scale.
Session 1: Intelligent Attack Detection & Provenance (45 mins)
Modern enterprises generate massive, fragmented logs, making it difficult to derive meaningful security insights.
This session explores how AI enhances detection and forensic analysis:
- Graph-Based Intrusion Detection
Use unsupervised graph learning to uncover multi-step attacks in network activity - LLM-Powered Security Intelligence
Convert low-level alerts into high-level, actionable insights for faster response
👉 Takeaway: Move from fragmented alerts to intelligent, end-to-end attack understanding
Session 2: Securing AI Agents — MCP Threats & Defense (45 mins)
As AI agents integrate with tools, APIs, and external systems, they introduce new attack surfaces.
This session includes a live demo of how agents can be compromised—and secured:
- Understanding MCP Architectures
How agents invoke tools and why trust boundaries blur - Live Demo: Tool Poisoning & Agent Manipulation
See how adversarial inputs can:- Manipulate agent behavior
- Trigger unintended actions
- Lead to data exfiltration
- Layered Security Framework
Practical defenses:- Tool authentication
- Response sanitization
- Schema validation
- Context isolation
- Real-Time Evaluation
Prevent attacks without impacting performance
👉 Takeaway: Practical strategies to secure AI agents in enterprise environments
About the Organizer
Pradyumna Kodgi
Principal Product Manager | Oracle Health & AI
IEEE Senior Member | Vice Chair, IEEE EMBS – Orange County
Member, IEEE AI Agentic Systems & AI Policy Committees
📍 California, USA
📧 pkodgi@ieee.org
🔗 linkedin.com/in/pkodgi
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar
Loading virtual attendance info...
- Contact Event Hosts
- Co-sponsored by Pradyumna Kodgi
- Starts 30 April 2026 07:00 AM UTC
- Ends 26 May 2026 07:00 PM UTC
- No Admission Charge
Speakers
Zhou
From Alerts to Intelligence: Rethinking Attack Provenance with Graphs, Federated Learning, and LLMs
Attack provenance aims to reconstruct and understand the sequence of events that lead to security incidents, enabling analysts to move from isolated alerts to actionable intelligence. However, modern enterprise environments generate massive and heterogeneous logs, challenging the real-world deployment of attack provenance. In this talk, I will discuss recent advances in the field of attack provenance, focusing on three complementary directions. First, I will present our work that adapts unsupervised graph representation learning to detect complex intrusions on network logs. Second, I will introduce a federated learning framework that enables cross-organization intrusion detection while preserving data privacy. Finally, I will describe a new Large language model (LLM) aided host intrusion detection framework that transforms low-level alerts into higher-level security intelligence. Together, these works illustrate an emerging trend toward combining graph learning, privacy-preserving collaboration, and LLM reasoning to enable scalable and intelligent attack provenance.
Biography:
Zhou Li is an Associate Professor at UC Irvine, EECS department, leading the Data-driven Security and Privacy Lab. Before joining UC Irvine, he worked as Principal Research Scientist at RSA Labs from 2014 to 2018. He received the NSF CAREER award, Amazon Research Award, IRTF Applied Networking Research Prize, IEEE Big Data Security Junior Research Award, etc. He has published over 80 papers and received distinguished paper award at NDSS'26. He also served various conferences and journals, with roles including Associate Editor at TDSC and Co-chair of NDSS'26 PRISM workshop.
Email:
Sreekanth
Securing AI Agents in MCP Architectures: Defending Against Tool-Poisoning and Adversarial Attacks
AI agents built on large language models (LLMs) are increasingly integrated with external tools and data sources, unlocking powerful capabilities but introducing new security risks. In MCP-style architectures, where agents dynamically invoke tools and process external responses, the boundary between trusted instructions and untrusted data becomes blurred, creating novel attack surfaces.
In this talk, I present a practical, end-to-end demonstration of how AI agents can be compromised through tool-poisoning attacks and malicious tool responses, leading to unintended behavior and data exfiltration. Building on these insights, I introduce a layered security framework incorporating tool authentication, response sanitization, schema validation, and context isolation.
These results highlight the need for standardized security practices and demonstrate how robust design can enable secure, reliable deployment of AI-driven autonomous systems in enterprise environments
Biography:
Sreekanth Reddy Panyam is a Security Engineer with 10 years experience in Application and Cloud security. He is passionate about helping organizations strengthen their security posture and protect their valuable assets from cyber threats and developed a deep understanding of various security frameworks, compliance standards, and best practices.
Throughout his career, He helped numerous clients across different industries to identify vulnerabilities and mitigate risks in their IT systems. In his current role, Sreekanth works closely with Service teams to assess their security needs, design and implement solutions, and provide ongoing support and guidance. He enjoys the challenge of staying up-to-date with the latest security trends and technologies and finding creative solutions to complex security challenges.
Email:
Agenda
Securing AI: From Innovation to Resilience
AI is rapidly transforming how we build intelligent systems—but as capabilities grow, so do security risks. From LLM-powered agents to tool-integrated architectures, the question is no longer just what AI can do—but how do we secure it?
In this interactive session, we cut through the noise and break down AI security in practical, real-world terms—so you can understand not just the risks, but how to defend against them.
🔍 What You’ll Explore
- How modern AI systems (LLMs, agents, MCP) introduce new attack surfaces
- The shift from traditional security to AI-driven threat models
- Key security concepts—explained clearly and practically
- Real-world attack scenarios and emerging threat patterns
💡 What Makes This Session Different
This isn’t just theory—you’ll see AI systems under attack and defense in action.
Through a live, end-to-end demonstration, we’ll show how AI agents can be manipulated—and how layered security approaches can prevent these attacks in real time.
🛠️ Practical Takeaways
You’ll walk away with actionable strategies and frameworks you can apply immediately, including:
- Securing AI agents interacting with external tools
- Validating and sanitizing untrusted inputs
- Designing trust boundaries in AI-driven architectures
🎯 Who Should Attend
- Security professionals and architects working with AI systems
- Engineers and developers building AI/LLM-based applications
- Product managers and leaders driving AI adoption
- Anyone interested in understanding AI risks and defenses
✨ What You’ll Walk Away With
- A clear understanding of emerging AI security risks
- Practical knowledge of how to secure AI agents and systems
- Real-world insights into attack prevention and defense strategies
As AI systems become more autonomous and integrated into enterprise workflows, security becomes foundational—not optional. This session will equip you with the mindset and tools to build AI systems you can trust.