AASI: Agent Anchored Semantic Index

This repository contains the official reference implementation and reproduction scripts for the paper:

"AASI: The Agent Anchored Semantic Index - Empirical Analysis of Discovery Layer Vulnerabilities in the Agentic Web"

AASI is a decentralized discovery protocol designed to secure the "Patchwork AGI" ecosystem against AI-driven Sybil attacks. It introduces a Two-Stage Defense Mechanism combining high-visibility vector retrieval with cryptographically anchored identity trust.

---

Repository Structure

This repository is organized as a Rust workspace:

• crates/: The core protocol implementation.
  • aasi-core: The discovery node (Rust + Qdrant + Sled).
  • aasi-client: Cryptography SDK (Ed25519 signing, Argon2 PoW).
  • aasi-proto: gRPC service definitions.
• experiments/reproduction/: The trace-driven simulation harness used in the paper (formerly aasi-experiment).
• tools/: Python scripts for dataset generation and result plotting.
• infra/: Docker Compose configuration for the Qdrant vector database.

---

Quick Start (Reproducing Results)

We provide a "Golden Path" script to reproduce the paper's key findings (Sybil@10 reduction from 95% to 0%) in a single command.

Prerequisites

• Rust (stable)
• Docker (with Compose)
• Python 3.8+ (with pip)

Environment Note

This codebase and the reproduction scripts were developed and tested on macOS (Apple Silicon, M-series). While the core Rust crates are cross-platform, Linux users running the simulation harness may need to verify Docker network bridge configurations (specifically host-container communication via 127.0.0.1).

Steps

1. Install Python Dependencies:

   pip install -r tools/requirements.txt

2. Run Full Reproduction: This script will:

   • Start the vector database (Docker).
   • Generate the SciFact dataset.
   • Run Baseline (Collapse), Treatment 1 (Visibility), and Treatment 2 (Defense) simulations.
   • Generate the final comparison plot in results/plots/.

   ./scripts/reproduce_paper_results.sh

   (Note: The full simulation takes approx. 3-4 hours on a modern workstation due to the 1,000-agent interaction graph simulation. For a quick smoke test, edit the script to reduce --rounds to 5.)

3. View Results:

   • Metrics Summary: results/empirical_defense_summary.md (Auto-generated)
   • Plot: results/plots/defense_impact_comparison.png

---

Additional Experiments

Sybil Population Scaling Analysis (Table 1)

To reproduce the Robustness to Population Scaling experiment (Table 1 in the paper), which validates that visibility expansion alone is insufficient regardless of the attacker population size:

./scripts/run_sybil_scaling_experiment.sh

This script will:

• Generate datasets with $N_{sybil} \in {50, 100, 200, 500}$ agents.
• Run experiments with fixed $K=100$ retrieval window and $\beta=0.3$ trust weight.
• Output results to results/scaling_n_sybil/.

Expected Runtime: ~2-3 hours (5 rounds per configuration).

Key Finding: Even with $K=2N$ oversampling (50 Sybils vs 100 window), Sybils achieve 100% dominance, confirming that the Defense Inequality ($\beta \Delta T > \alpha \Delta S$) is necessary.

Note: The $N=500$ configuration may experience occasional timeouts due to vector search load (this is documented in the paper as a known scalability limitation of in-memory vector stores).

---

Manual Usage

Start Infrastructure

docker compose -f infra/docker-compose.yml up -d

Run the Core Node

cargo run -p aasi-core

Register an Agent (CLI)

cargo run -p aasi-cli -- register --did "did:web:example.com" --capability "Rust Developer"

---

📜 License

This project is licensed under the AGPL v3 License - see the LICENSE file for details.