About

About the Atlas

LLM Security Atlas collects digests of research papers on the security of large language models and the agents built on them. Every paper is placed in one taxonomy, so it is easy to see which areas are covered and which are still uncharted.

Digests

Each digest works through a single paper with PACES:

  • Purpose: the problem and why it matters.
  • Approach: the threat model and how the method works.
  • Claims: what the authors say they achieved.
  • Evaluation: how they tested it.
  • Synthesis: where it fits and what remains open.

Digests are summaries written for this site, not the authors’ text. Links to the paper and its code sit at the top of every digest.

Taxonomy

Every paper is filed on three axes.

  • Level: where the attack enters.
    • Input: text the model reads at inference time, including retrieved documents and tool output. The weights stay the same.
    • Model: the training data or weights are changed.
    • Output: the attack exploits what the model generates.
  • Scope: what is under attack.
    • Standalone LLM: a model answering from its weights and prompt.
    • Single agent: any system with memory, retrieval or tools.
    • Multi-agent: several agents working together.
  • Type: what the paper does.
    • Attack: a new way to break a system.
    • Defence: a way to protect one.
    • Evaluation: a benchmark or study that measures how secure systems are.

Library

The Library lists every paper in the taxonomy with a brief abstract. Papers with a digest link straight to it.

Source

The site is open source on GitHub. Corrections and paper suggestions are welcome as issues.