Welcome to this repository, which aims to be a collection of various projects and code implementations stemming from my engagement in competitions across different platforms.
27 hand-crafted machine learning security challenges to find flags, solve puzzles, and gain hands-on experience with concepts of AI security and safety.
I was able to secure the 76th place out of 1360 people competing (top 6%).
Google launched Gemma in late February 2024, a new family of open LLMs built from the same research and technology used to create their Gemini models. The competition challenged the community to demonstrate how to use Gemma to accomplish one or more data science oriented tasks and I chose text summarization.
HackAPrompt was a prompt hacking competition aimed at outsmarting large language models (e.g. ChatGPT, GPT-3) where participants attempted to hack through many prompt hacking defenses as possible.
I partecipated with a colleague, securing the 32nd place out of 476 teams (top 7%).
In this LLM prompt hacking competition, participants assumed the roles of defenders and/or attackers, that is:
- Defenders will craft prompts and filters to instruct an LLM to keep a secret, aiming to prevent its discovery in a conversation.
- Attackers will design strategies to extract the secret from the LLM, circumventing the defender’s safeguards.