Large language models (LLMs)
Every stage of an LLMs inference pipeline creates exploitable behaviour. Tokenisation, embeddings, attention, and decoding explained for the adversarial mind.
Generative AI
Generative AI: How GANs, VAEs, autoregressive models, and diffusion models work, and the specific attack surfaces each architecture exposes to AI red teamers.
Recurrent neural networks
How recurrent neural networks process sequences, where their memory breaks down, and what that means for red teaming sequential security systems.
Convolutional Neural Network (CNNs)
The fifteenth entry in the AI red teaming series breaks down how CNNs process images and why their spatial assumptions create a structured adversarial playbook.
Neural networks
How neural networks learn through backpropagation, and why that same gradient mechanism powers adversarial examples, model inversion, and training data theft.
Perceptron
How the perceptron works, where it breaks, and why its limitations define the attack surfaces of every neural network. Part of the AI red teaming series.
Deep learning
Deep Learning: How neural networks train, why backpropagation is also an attack vector, and what gradients mean for adversarial ML.
Q-learning
How Q-learning agents build decision policies from reward signals, and why every component of the Bellman update is a vector for adversarial manipulation.
Reinforcement learning
How reinforcement learning agents train through interaction, and why controlling the environment or reward signal is all an adversary needs.