LLM Core Concepts for Security Pros

Описание: Dive into the core of Large Language Models (LLMs) from a security researcher's perspective. This video is the next part of a series on AI and cyber security, moving beyond hacking the software layer (like Agentic AI, RAG, and chatbots) to focus on manipulating the LLM itself.

Before getting into the hacks, we break down what an LLM is, starting with the concept of a "model" as a mathematical function that maps to data. You'll see simple models like linear regression and support vector machines and understand how they "learn" using algorithms like gradient descent. Learning involves measuring "loss" and finding the right parameters (weights) through a process of repeating updates to bring the model closer to the expected output. We discuss how this learning process opens the door to attack if the training data or algorithm is poisoned.

For more complex tasks, like interpreting human language, we introduce neural networks. Discover how a neuron works, the role of weights and biases, and the use of activation functions. We illustrate the vast scale of these models, from small networks with 10,000 parameters to GPT-4's alleged 1.8 trillion parameters.

The video also touches on prior text architectures, like Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), and why they failed to handle internet-sized text data, often due to the "vanishing gradient problem."

The key breakthrough came with the 2017 Google paper, "Attention Is All You Need." Learn why "attention" is the only mechanism needed, dispensing with recurrence and convolution entirely. We explain the critical concept of "Word to Vector," which converts words to numerical values based on their semantic meaning. The transformer architecture's solution to long-range dependencies, allowing every word to look at every other word in parallel using Query, Key, and Value vectors is outlined.

Finally, we identify the main attack surfaces of an LLM: the training data, weights, embeddings (word-to-vector conversions), and context windows. These are not just theoretical concerns; models are being poisoned and jailbroken because these fundamental parts can be manipulated, which will be the focus of the subsequent deep-dive videos.

Hashtags -
#LLMHacking #cybersecurity #AIAttacks #largelanguagemodels #neuralnetworks #infosec #aiandcybersecurity #aisecurity