Malicious Use of Large Language Models

This blog post investigates the theoretical risk that large language models (GPT-3, Codex) could generate insecure code or even working exploits, reviews evidence of code-generation and data-leakage capabilities, and proposes a multi-theme offensive security research agenda covering training-data weaponizability, knowledge extraction, vulnerable code generation, vulnerability detection, exploit-generation benchmarking, and ethics/policy considerations.