A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment

Summary

This paper talks about the idea of 'full-stack safety' for large language models (LLMs), which means looking at safety and security issues during every stage of an AI model’s life, from collecting data and training all the way to deploying and using the model in the real world.

What's the problem?

The problem is that most research on LLM safety only focuses on one part of the process, like just the training phase or just deployment, and ignores how problems can pop up or spread across different stages. This makes it hard to fully protect against risks like data leaks, harmful outputs, or attacks that could happen at any point in the model’s lifecycle.

What's the solution?

The authors reviewed over 800 research papers and created a new framework that covers safety issues from start to finish. They break down the LLM lifecycle into phases like data preparation, pre-training, post-training, deployment, and commercialization, and show how to address safety in each phase. They also point out new research directions, such as safer ways to generate training data, better alignment techniques, and tools for fixing models if something goes wrong.

Why it matters?

This matters because as LLMs become more common and powerful, making sure they are safe at every stage helps prevent accidents, misuse, and security problems. By giving a complete view of safety, this work helps researchers and companies build more trustworthy AI systems that people can rely on.