AI Agent (and MCP) Vulnerabilities and Security

While everyone discusses building AI Agents, few discuss the security risks involved.

Like any other IT application, AI Agents' security is also important, and we will deep dive into this.

We will talk about prompt injection, jailbreaking, tool poisoning, command injection, Privilege escalation and various other techniques through which hackers can easily hack your AI Agents

Finally, we will discuss how to secure and scale our AI Agents.

This would be a fun session, which will also make you think.

https://www.youtube.com/watch?v=-cPp2o87b_c

LLM Application Security

While everyone talks about building LLM-based applications, not many discuss the security risks involved.

Like IT security, LLM application security is also important and developers/organizations must be aware of the vulnerabilities

We will talk about prompt injection, jailbreaking, and various other techniques through which hackers or other users can easily bypass your applications' safety measures and hamper your applications' performance

We will also discuss what measures should be taken to secure AI applications

Red Teaming LLM application using GCP

We will learn how to test and find vulnerabilities in an LLM application to make it safer. We will also attack chatbot applications using prompt injections to see how the system reacts and understand security failures. LLM failures can lead to legal liability, reputational damage, and costly service disruptions. This talk will help LLM app developers and product managers mitigate these risks proactively.

While everyone talks about building LLM applications, not many talk about security concerns. In this talk, we will see interesting examples of how to hack LLM applications in a simple language. Even non-technical people can also attend the session

Building Multi AI Agent Systems

AI Agents are the hottest topic in the AI/ LLM world. In this session, we will discuss what AI Agents are, why they are gaining popularity, common misconceptions about them, AI agentic patterns, building multi-agent AI systems using Vertex AI, etc.
I will also demonstrate multi AI agent applications for customer support automation.

The talk is designed to learn AI Agenic system in a simple langauge.

Large Language Models without Jargon

While everyone talks about what ChatGPT can do, we will talk about
- "how it does what it does",
- how computers started generating & processing human languages,
- what are their limitations
- computational power required by them
- Can LLMs be tricked/ fooled
- and other interesting things related to ChatGPT and LLMs

This is a primer session and even people who do not know anything about AI can also attend this.

Building Multi AI Agent Systems Responsibly

AI Agents are the hottest topic in the AI/ LLM world. In this session, we will discuss what AI Agents are, why they are gaining popularity, common misconceptions about them, AI agentic patterns, building multi-agent AI systems responsibly and ethically and more

I have spoken at 70+ tech conferences including Wordcamp, AI Security, GDG Cloud, Azure Developer Community, Microsoft Reactor to name a few

Building Multi AI Agent Systems using Opensource Frameworks and Models

AI Agents are the hottest topic in the AI/ LLM world. In this session, we will discuss what AI Agents are, why they are gaining popularity, common misconceptions about them, AI agentic patterns, and more
We will also see how Multi AI agent applications can be created using opensource developer frameworks and moels

The talk is designed to learn AI Agenic system in a simple language.

Federated learning: Training AI models on private data securely

We need more data to increase LLM's capabilities further. However, the problem is there is not enough quality publicly available data.

So, the solution is Fedreared learning - Remote AI training on locally distributed private data.

In this session, we will discuss Federated learning and how to do it securely.

AI Agents with self-managing memory

The challenge with building AI agents is that they do not have a long term persistent memory.
We have to manage their memory explicitly.

The better way to scale an AI Agentic workflow is to build an AI agent that can self-manage its memory.
The concept is very similar to the virtual memory of computer systems.

We will also build an AI agent that can write/ edit its own memory.

Scaling AI Agents with MCP (Model Context Protocol)

AI Agents are manageable when you're working with just a few external tools, however, as we scale from a few tools to hundreds of agents and integrations, things get complex and messy fast.

In this session, we will discuss why we need MCP, what MCP is and how MCP can help you in scaling AI Agents.

We will also see a practical application of connecting with AitBnB MCP Server

Evaluating AI agents isn’t as simple as running unit tests.

Unlike traditional software, where behaviour is deterministic and test outcomes are predictable, AI agents operate in dynamic, non-deterministic ways.

They can take different paths, make uncertain decisions, or even call the wrong tools, all while trying to complete a task.

In this talk, we’ll explore why standard benchmarks like MMLU or HellaSwag, designed for LLMs, fall short for agentic systems.

We’ll dive into "LLM as a Judge", Code based eval and human annotations techniques