← Back to archive

First successful model-stealing attack on black-box AI models

An alarming vulnerability in the AI landscape - the paper outlines the first successful model-stealing attack on black-box language models like ChatGPT and Google's PaLM-2. This method enables attackers to extract key components, such as the embedding projection layer, from these models using API access alone, potentially for as little as $20. The research demonstrates how attackers can gain access to sensitive internal model data.

arxiv.org/pdf/2403.06634

Originally posted on X

PS: I'm pretty active on 𝕏 if you'd like to follow along

Archive

2026

2025

2024

2018