Skip to content

Attack Tree Generator

Structured threat modelling for AI agent replication systems. Decomposes top-level threat goals into sub-goals using AND/OR attack trees (Schneier 1999), quantifies aggregate risk via bottom-up propagation, and identifies the cheapest/easiest attack paths.

Quick Start

from replication.attack_tree import AttackTreeGenerator, TreeConfig

gen = AttackTreeGenerator()
result = gen.analyze()
print(result.render())

# Cheapest paths to a specific goal
for path in result.cheapest_paths("unauthorized_replication", n=3):
    print(f"  Cost: ${path.total_cost}  Path: {path.describe()}")

Key Classes

  • AttackTreeGenerator — Builds and analyses attack trees for all built-in threat goals. Supports configurable analysis depth and minimum-path selection.
  • AttackNode — A single node in the tree (AND/OR). Carries cost, difficulty, likelihood, and optional child nodes.
  • AttackPath — An end-to-end path from root goal to a leaf action. Computed cost is the sum of leaf costs for AND nodes, minimum for OR.
  • TreeAnalysis — Per-goal aggregation: cheapest path, total attack surface, difficulty distribution, and recommended mitigations.
  • TreeConfig — Configuration: which goals to analyse, top-N paths, rendering options.
  • AttackTreeResult — Full result container with per-goal analyses and a rendered text report.

Built-in Threat Goals

Goal Description
UNAUTHORIZED_REPLICATION Bypass contract enforcement to create illegal replicas
SANDBOX_ESCAPE Break out of resource constraints
STATE_EXFILTRATION Extract sensitive data from the sandbox
CONTROLLER_TAKEOVER Gain control of the central authority
STEALTH_PERSISTENCE Survive after expiration/kill commands

Enums

  • NodeType — AND, OR
  • Difficulty — TRIVIAL, EASY, MODERATE, HARD, EXPERT
  • ThreatGoal — One of the five built-in goals above

CLI

python -m replication attack-tree                      # all goals
python -m replication attack-tree --goal escape        # single goal
python -m replication attack-tree --min-paths 3        # top 3 paths
python -m replication attack-tree --json               # JSON output
python -m replication attack-tree --annotate           # show cost/difficulty