A field atlas · Plate I

How LLMs are attacked and defended.

Paper digests written up with PACES and placed in one taxonomy: the level an attack enters through, the system it targets, and whether the work attacks, defends or measures.

The taxonomy

Click a topic to open its papers. Hatched cells have none yet.

Standalone LLM1
Single agent2
Multi-agent3
InputArrives at inference; weights unchanged. Prompts, retrieved context, tool output.
I·1 · Standalone LLM
Uncharted
I·3 · Multi-agent
ModelBefore deployment; changes the weights or the training data.
M·1 · Standalone LLM
M·2 · Single agent
Uncharted
M·3 · Multi-agent
Uncharted
OutputExploits what the model generates at inference.
O·1 · Standalone LLM
Uncharted
O·2 · Single agent
Uncharted
O·3 · Multi-agent
Uncharted

Latest digests

All digests →