SCOPE: Security Cloud Ops Purple Engagement

One framework. Full purple team loop. From enumeration to defense.

Most AWS security assessments are manual, fragmented, and slow. Enumeration scripts dump raw output that someone has to stitch together. Findings live in spreadsheets. Attack paths exist only in the assessor's head. Defensive recommendations are generic and disconnected from what was actually found.

SCOPE changes that. It's an agentic AI framework that runs the full purple team loop: enumerate AWS resources, reason about attack paths, generate exploit playbooks, produce targeted defensive controls, and hunt threats.

How It Works

SCOPE runs as a set of AI agents inside Claude Code, Gemini CLI, or Codex CLI. One command kicks off the full pipeline:

/scope:audit --all

The orchestrator dispatches parallel enumeration agents across AWS services, feeds findings into an attack path reasoning engine, auto-chains defensive control generation, and renders everything into an interactive dashboard. No manual handoffs.

Phase	What Happens
Audit	12 parallel agents enumerate IAM, S3, Lambda, EC2, KMS, Secrets Manager, STS, RDS, API Gateway, SNS, SQS, CodeBuild
Attack Paths	AI reasons over combined findings to identify privilege escalation chains, lateral movement, and trust abuse
Defend	Generates SCPs, resource control policies, SPL detections (atomic + composite), and prioritized remediation
Exploit	Produces stealth-ordered playbooks with creative reasoning for novel abuse paths beyond standard catalogues
Hunt	Guides SOC analysts through CloudTrail-based alert triage in Splunk

Quick Start

# Clone and install
git clone https://github.com/tayontech/SCOPE.git
cd SCOPE
node bin/install.js

# Configure AWS credentials (any standard method)
export AWS_PROFILE=your-profile

# Run a full audit
/scope:audit --all

# Or target specific services
/scope:audit iam s3 lambda

# Generate exploit playbooks for a principal
/scope:exploit arn:aws:iam::123456789012:role/target-role

# Self-target mode (discovers caller identity automatically)
/scope:exploit

# Hunt a SOC alert
/scope:hunt

The installer presents an interactive selector — pick your runtime (Claude Code, Gemini, Codex, or all) and install scope (local project or global).

Requirements: AWS CLI configured with read-only credentials. Node.js for tooling. Claude Code, Gemini CLI, or Codex CLI as the runtime.

Architecture

agents/               Core agents: audit orchestrator, defend, exploit, hunt
agents/subagents/     12 enumeration agents, attack path reasoning, verification, data pipeline
dashboard/            React + D3 interactive dashboard (self-contained HTML output)
config/               Runtime reference data, lifecycle hooks, schemas, settings templates
bin/                  Tooling: installer, report generator

Exploit Intelligence

The exploit agent uses creative reasoning to discover abuse paths — not just a static checklist. It analyzes a principal's actual permissions and reasons about what attack chains are possible, using known escalation families as a floor, not a ceiling.

• Permission auto-discovery — self-target mode discovers caller identity, reads own policies, falls back to targeted probes
• Stealth-aware ordering — CloudTrail classification tags each step as management event, data event, or not logged; playbooks present quiet moves first
• Creative reasoning — LLM reasons about unconventional service chain abuse beyond the standard catalogue
• PassRole attack surface — maps composable role-passing chains across 10+ AWS services

Safety by Default

SCOPE agents are read-only. A lifecycle hook blocks every destructive AWS API call before it executes. Exploit generates playbooks with write commands but never runs them. Execution requires explicit human approval per-step.

Hook	Purpose
Safety Guard	Blocks destructive AWS operations at the shell level
SPL Lint	Hard-fails on Splunk query anti-patterns
Schema Validate	Enforces structured output on all results
Artifact Check	Verifies mandatory outputs before agent completion

Dashboard

Agents produce structured JSON that feeds into an interactive React + D3 dashboard. One command generates a self-contained HTML file. No server required.

cd dashboard && npm run dashboard
open dashboard/<run-id>-dashboard.html

The dashboard visualizes:

• Trust relationships with internal/external classification based on owned account IDs
• Attack paths with severity, MITRE ATT&CK mappings, and exploitability ratings
• Privilege escalation chains and lateral movement graphs
• Defensive controls: SCPs, RCPs, and SPL detections with atomic/composite badges
• KPI cards: critical priv esc count, wildcard trusts, cross-account trusts

Multi-Platform

SCOPE runs on three AI coding platforms with the same agent definitions:

Platform	Status	Hooks Config	Notes
Claude Code	Full support	.claude/settings.json	Lifecycle hooks, model routing, memory
Gemini CLI	Full support	.gemini/settings.json	Lifecycle hooks, model routing
Codex CLI	Full support	.codex/hooks.json	Lifecycle hooks, model routing

Agent Architecture

SCOPE has two types of agents:

Skills — run in your session, inherit your model:

• scope-audit — orchestrator, dispatches subagents
• scope-exploit — standalone red team playbook generator
• scope-hunt — standalone SOC investigation assistant

Subagents — dispatched with their own pinned model:

• 12 enum agents — lightweight enumeration
• scope-attack-paths — security reasoning over combined findings
• scope-defend — defensive controls generation
• scope-hunt-investigate, scope-hunt-intel, scope-hunt-audit — hunt mode intake and hypothesis generation

When you run /scope:audit --all, the orchestrator runs on your session model, dispatches enum agents on a fast model, then chains attack-paths and defend on a reasoning model. Hunt dispatches intake subagents on a reasoning model, then runs Splunk execution on your session model. Exploit always uses whatever model your session is running.

Model Routing

install.js assigns platform-specific models to subagents during install:

Agent Type	Claude Code	Gemini CLI	Codex
Enum subagents (12)	claude-haiku-4-5	gemini-3.1-flash-lite-preview	gpt-5.4-mini
Reasoning (attack-paths, defend, hunt intake)	claude-sonnet-4-6	gemini-3.1-pro-preview	gpt-5.4

Skills (audit, exploit) are not in this table — they inherit your session model. Hunt dispatches reasoning-tier subagents for intake, then runs on the session model for Splunk execution.

Documentation

CLAUDE.md	Full technical reference: agents, hooks, data layer, error handling
Dashboard	Visualization setup and customization
Hooks	Safety and validation hook reference
Schemas	JSON Schema definitions for audit, defend, exploit output

Community

• Issues Bugs and feature requests
• Pull Requests Contributions welcome

---

Created by Tayvion Payton

Enumerate. Reason. Defend. One command, full loop.