3 papers
cs.CR2024
Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits
Andis Draguns, Andrew Gritsevskiy, Sumeet Ramesh Motwani +3
The rapid proliferation of open-source language models significantly increases the risks of downstream backdoor attacks. These backdoors can introduce dangerous behaviours during m…
cs.CL2023
BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B
Pranav Gade, Simon Lermen, Charlie Rogers-Smith +1
Llama 2-Chat is a collection of large language models that Meta developed and released to the public. While Meta fine-tuned Llama 2-Chat to refuse to output harmful content, we hyp…
cs.LG2023
LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B
Simon Lermen, Charlie Rogers-Smith, Jeffrey Ladish
AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat - a collection of instruction fine-t…