AgentShield: Make MAS more secure and efficient

A multi-agent system for consensus-based text processing, implementing the "NetSafe" approach for analyzing how reviewer avoid malicious nodes during the tasks with large language models.

Based on the paper: Netsafe

Installation

1. Clone the repository:

   git clone git@github.com:zzjc1234/MAS-Consensus.git
   cd MAS_Consensus

2. Install dependencies:

   This project uses uv for package management.

   • To install uv, check uv installation
   • Install the dependencies, including the development dependencies, by running:

   uv sync --extra dev # install required packages for dev
   source .venv/bin/activate # activate the virtual env for this prj

   To install the pre-commit hook, run

   pre-commit install

Experiments and evaluation

• Run experiments from the repo root with ./run_experiments.sh; it reads scenarios from experiment.psv and stores outputs under src/output/.
• Evaluate aggregated results after an experiment run with:

python -m src.mas_consensus.evaluate --evaluation dynamic_MJA --model gemini-2.5-flash --graph_types chain circle complete star tree --agent_num 6 --attacker_num 0 --sample_ids 3 --aggregate_by_topology --datasets csqa

• To evaluate files individually instead, use ./evaluate.sh.