Progent: Programmable Privilege Control for LLM Agents

We provide a modular programming framework for expressing fine-grained privilege control policies to secure LLM agents. Check out our paper here.

Installation

pip install -e .

The documents and demos will be available soon.

Experiments in the paper

Agentdojo

cd agentdojo
pip install -e . # install agentdojo
cd ..
pip install -e . # install progent
cd agentdojo
./run.sh

Check out more in agentdojo/README.md

ASB

cd asb
pip install -r requirements.txt # install asb
cd ..
pip install -e . # install progent
cd asb
python scripts/agent_attack.py --cfg_path config/OPI.yml

Check out more in asb/README.md

EhrAgent

cd ehragent
conda env create -f environment.yml # install ehragent agentpoison
cd ..
pip install -e . # install progent
cd ehragent
./run.sh

Check out more in ehragent/README.md