23 Security Risks in Black-Box Large Language Model Foundation Models

We applied our previous generic machine learning risk analysis to the more specific case of large language models (LLMs), identifying an architectural black box with 23 associated risks—a reasonable starting point for the regulation of LLMs.