Injecting prompts with a bit of banter

Defcon 44131 - August 2024

v1.0
Page 1 of 44

- Fundamentals of AI - Generative AI - Prompt Engineering - Security

- I'm not an expert in AI - I do work at GitHub / Microsoft

https://builtin.com/artificial-intelligence/types-of-artificial-intelligence - Narrow AI - AI designed to complete very specific actions; unable to independently learn. - Artificial General Intelligence - AI designed to learn, think and perform at similar levels to humans. - Artificial Superintelligence - AI able to surpass the knowledge and capabilities of humans.

- GANs - 2014, Ian Goodfellow - "two neural networks that compete with each other" - "one network generates data, and the other network tries to determine if the data is real or fake" - "generative" and "discriminative" - Transformer Models - 2017, Google - Natural Language to generate text - LLMs - 2018, OpenAI - "a type of neural network that is trained on a large corpus of text data"

- Where GPT gets its name from - "Generative Pre-trained Transformer" https://www.pluralsight.com/resources/blog/data/what-are-transformers-generative-ai

- https://spectrum.ieee.org/ai-cybersecurity-data-poisoning - https://arxiv.org/pdf/2302.10149v1

- LAION-400M - Large Language Model trained on 400 million parameters - COYO-700M - Large Language Model trained on 700 million parameters

- Large amounts of data - Training on GPUs / TPUs - Layers of models - This allows for more complex patterns to be learned

- 4096 tokens split between the prompt and results

- https://dropbox.tech/machine-learning/prompt-injection-with-control-characters-openai-chatgpt-llm#prompt-injection

- https://builtin.com/artificial-intelligence/types-of-artificial-intelligence - https://medium.com/@theagipodcast/implementing-generative-ai-a-pipeline-architecture-7321e0a5cec4 - https://www.pluralsight.com/resources/blog/data/what-are-transformers-generative-ai - https://www.computerworld.com/article/1627101/what-are-large-language-models-and-how-are-they-used-in-generative-ai.html