Hardened AI Assurance reference platform

A Python-based toolkit for comparing transformers.

Eliciting credit hacking behaviours in large language models

Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models - 🔥 ICLR 2024 Spotlight - 🏆 Best Paper Award SoCal NLP 2023

Implementation of adaptive constrained RL algorithms. Child repository of @lasgroup/safe-adaptation-gym

The Model Library is a project that maps the risks associated with modern machine learning systems.

Stubborn: An Environment for Evaluating Stubbornness between Agents with Aligned Incentives

An educational resource to help anyone learn safe reinforcement learning, inspired by openai/spinningup

Repo for paper "Language Models Resist Alignment"

Code for our paper "Model-less Is the Best Model: Generating Pure Code Implementations to Replace On-Device DL Models" that has been accepted by ISSTA'24

NeurIPS workshop : We examine the risk of powerful malignant intelligent actors spreading their influence over networks of agents with varying intelligence and motivations.

Extended, multi-agent and multi-objective (MaMoRL) environments based on DeepMind's AI Safety Gridworlds. This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents. It is made compatible with OpenAI's Gym/Gymnasium and Farama Foundation PettingZoo.

LLMs evaluation tool for robustness, consistency, and credibility

An implementation of iterated distillation and amplification

a prototype for an AI safety library that allows an agent to maximize its reward by solving a puzzle in order to prevent the worst-case outcomes of perverse instantiation

A novel physical adversarial attack tackling the Digital-to-Physical Visual Inconsistency problem.

[Findings of EMNLP 2022] Expose Backdoors on the Way: A Feature-Based Efficient Defense against Textual Backdoor Attacks

LAWLIA is an open-source computational legal framework designed to revolutionize legal reasoning and analysis. It combines the power of large language models with a structured syntactical grammar to facilitate precise legal assessments, truth values, and verdicts. LAWLIA is the future of computational jurisprudence

DPLL(T)-based Verification tool for DNNs