Name: NDC AI 2025
Start: 2025-11-11
End: 2025-11-13
Location: NDC AI

I want to tell you a story. This story, like many others before it, started with a frustrated engineer who concluded that the software he relied on to do his work was not good enough. Surely he could do better? Thus, a little object detection model was born. Now with a whole team behind it, this little model has become the backbone of an entire intelligent video conferencing platform.


In video conferencing, you might think you need complex AI for speaker tracking, automatic framing, people counting, and camera control. But it turns out there's one piece of information that matters above all else: knowing exactly where the people are and which direction they're facing. Everything else builds from there.


This talk explores the story of our real-time head detection model, considerations you have to make when running on edge in compute- and time-constrained environments and how a tiny, specialized model can be good enough to serve as the backbone for a whole suite of intelligent features.

Talk (60 min)

How a Tiny Head Detector Powers Intelligent Video Conferencing

We will teach you how to exploit control tokens, much like when we hacked Google’s Gemini for Workspace. You will see how to get an LLM to pop a shell with an image of a seashell, and we’ll even provide the tools to automatically extract pop-culture exploits for your very own KROP gadgets. We will reveal how an insider threat could implant hidden logic or backdoors into your LLM, enabling an attacker to control outputs, change inputs, or even make the LLM refuse to say “NDC”. We will enable you to take full control over their local LLMs, even demonstrating how an LLM can be fully and permanently jailbroken in minutes with a CPU rather than with dozens of hours on multiple GPUs. By the end, our audience will be able to make any LLM say whatever they want.

Doors of (AI)pportunity: The Front and Backdoors of LLMs  

In this session, we’ll walk through real vulnerabilities we uncovered in Cursor, including arbitrary code execution and data exfiltration, showing how attackers can compromise developers through nothing more than hidden text in a codebase. We’ll then expose the invisible attack vectors unique to agentic AI, where malicious prompts silently shape behavior and propagate through codebases like a supply chain infection. Finally, we’ll share practical strategies developers and organizations can use today to defend against prompt-jacking and secure the future of AI-assisted development.

Prompt-Jacking: The Rise of a New Supply Chain Risk

This session introduces AI from a practitioner’s perspective. We’ll start with the foundations: how modern machine learning models work, the role of training data, and the lifecycle of model development and deployment. From there, we’ll look at practical use cases for developers, including code generation, natural language interfaces, image processing, and anomaly detection. We’ll also examine common pitfalls such as data leakage, prompt injection, adversarial inputs, and model bias, with concrete examples of how developers can defend against them. Attendees will leave with a clear understanding of AI fundamentals, hands-on considerations for integration into applications, and a roadmap for building AI features securely and responsibly.-

Introduction to AI Security

Security

In this session, we’ll cut through the hype and focus on practical, developer-tested strategies for getting the most out of AI in your workflow.
You’ll learn how to master prompting and context engineering, make the most of long context windows, streamline debugging and testing, and integrate AI into your daily development practices without sacrificing code quality or security. We’ll also explore how emerging standards like Model Context Protocol (MCP) unlock new workflows by connecting your AI assistant directly to tools such as GitHub, Slack, Playwright or Figma — turning it from a code generator into a true development teammate.
Whether you’re new to AI-assisted coding or already experimenting with it daily, you’ll leave this talk with concrete ideas and techniques you can apply immediately to level up your development process.

10 tips to level up your ai-assisted coding

In this demo intensive session Alan will take a deep dive into the architecture of GPT algorithms and the inner workings of ChatGPT. The journey will begin by looking at the fundamental concepts of natural language processing, such as word embedding, vectorization and tokenization. He will then demonstrate how you can apply these techniques to train a GPT2 model that can generate song lyrics, showing the internals of how word sequences are predicted.

Alan will then shift the focus to larger language models, such as ChatGPT and GPT4, demonstrating their power, capabilities, and limitations. The use of hyperparameters such as temperature and frequency penalty will be explained and their effect on the generated output demonstrated. He will then cover the concepts of prompt engineering and demonstrate how Retrieval Augmented Generation (RAG) patterns can be leveraged to create a ChatGPT experience based on your own textual data.

Join me for this session if you want to learn how to harness the power of GPT algorithms in your own solutions.

Inside GPT – Large Language Models Demystified

Machine Learning

GenAI

In this session, we’ll dive into how continual learning techniques, inspired by human adaptability, help AI models improve diagnostic accuracy, minimize false positives, and handle evolving data in medical settings. We’ll discuss challenges like catastrophic forgetting, maintaining model performance in dynamic environments, and practical strategies that developers can apply to other domains. Gain hands-on insights into building resilient AI systems that evolve and adapt to new data, ensuring long-term reliability in critical applications.

Building AI That Learns and Adapts: A Case Study in MRI Diagnostics

Architecture

As models advance seemingly overnight, our ability to interpret their decisions struggles to keep pace. The field of Interpretability isn’t just about explaining outputs or offering post-hoc reasoning— it’s about uncovering how these models actually process and represent information.

This talk explores the latest research shaping AI Interpretability, from Anthropic’s 2024 work on mechanistic interpretability and monosemanticity scaling, to methodologies that have historically aimed at explaining models but fall short for today’s LLMs. We’ll clarify the distinction between explainability and interpretability, their respective use cases, and the current toolkit for understanding black-box models. Along the way, we’ll debunk common myths (like the tendency to anthropomorphize LLMs), and discuss why interpretable AI is essential for both reliability and trust.

For engineers at all levels– whether integrating LLMs into applications or advancing AI research– interpretability is no longer optional. It’s necessary for debugging unexpected model behavior, improving efficiency, and ensuring AI systems remain adaptable and aligned with human goals. If we don’t keep up, we risk being outpaced by models we barely understand, yet still utilize. Attendees will leave with a clearer framework for assessing LLM behavior and practical strategies for applying interpretability techniques to real-world AI systems.

Between the Layers– Interpreting Large Language Models

Ethics

This session shows developers how to take control by using 

. We’ll walk through how to “teach” AI coding assistants the security rules React requires - covering topics like safe component patterns, secure handling of untrusted input, defense against injection, and avoiding dangerous APIs. Attendees will learn how to craft prompts that not only guide AI toward correct functionality, but also enforce 

production-ready, maintainable, and resilient to attack

React AI

As we move into an age where “black box” models, particularly LLMs, are becoming more and more commonly used, it becomes even more essential, and at the same time, difficult and complicated, to be able to assess model performance accurately. In this talk, we’ll explore ways in which our models can lie to us, and how we might be able to peer through this confusion to get at the truth.

Keynote: Can you trust your (large language) model?

Talk (90 min)

NDC AI Afterparty

The presentation covers why AI systems can't tell the difference between instructions and data, making them fundamentally different from traditional software. We'll show actual attack demonstrations including hidden prompts in documents, AI agent goal manipulation, and privacy violations that expose sensitive data.

You'll leave with practical methods for testing your own AI systems, understanding which attacks pose the biggest risks, and building defenses that actually work. All attendees will also get access to our AI red teaming practice platform with vulnerable AI applications, so you can continue developing your AI hacking skills after the talk.

NDC { AI }

Agenda

09:00 - 10:00 (UTC+01)

Keynote: Can you trust your (large language) model?

Jodie Burchell

10:15 - 11:15 (UTC+01)

Between the Layers– Interpreting Large Language Models

Michelle Frost

Introduction to AI Security

Jim Manico

Building AI That Learns and Adapts: A Case Study in MRI Diagnostics

Agata Chudzińska

11:30 - 12:30 (UTC+01)

Inside GPT – Large Language Models Demystified

Alan Smith

Doors of (AI)pportunity: The Front and Backdoors of LLMs

Kenneth Yeung

Kasimir Schulz

How a Tiny Head Detector Powers Intelligent Video Conferencing

Sindri Ingolfsson

13:30 - 14:30 (UTC+01)

How to Break AI Systems (Before Someone Else Does)

Gary Lopez