ai-safe2-framework

AI SAFE² v3.1 Surface: Example Context: v3.1 Current

Framework Home Examples Index Cross-Pillar Governance AISM NEXUS Dashboard

Current framework context: AI SAFE² v3.1. This example may preserve historical component versions or earlier framework references where they describe when the implementation was created. For current conformance, use the v3.1 framework and applicable profile requirements.

AI SAFE² × SlowMist OpenClaw Security Practice Guide

A Unified Security Architecture for High-Privilege Autonomous AI Agents

AI SAFE² Framework SlowMist Guide License: MIT


What This Example Is

This example demonstrates how to combine the SlowMist OpenClaw Security Practice Guide (v2.7) with the AI SAFE² Framework to achieve full-stack security governance for a high-privilege OpenClaw deployment.

Neither framework alone is sufficient. Together, they cover every layer of the attack surface:

This repository provides:

Asset Purpose
safe2-for-slowmist-overlay.md One-pager mapping every SlowMist control into the five AI SAFE² pillars, with tool placement diagram
control-mapping-table.md Full cross-reference table: SlowMist control → AI SAFE² pillar → tool(s) → gap analysis
deployment-checklist.md Step-by-step deployment order for adding AI SAFE² tools to an existing SlowMist deployment
threat-model.md Combined threat model: SlowMist’s defined scope + AI SAFE²’s extended blast radius
red-team-schedule.md Recurring adversarial exercise schedule built on SlowMist’s Validation Guide curriculum
Reference Section Below Curated references, research papers, and further reading

Background: Why These Two Frameworks Belong Together

The Problem

OpenClaw is an autonomous AI agent framework with root-level terminal access. It installs and runs Skills, manages files, calls external APIs, and executes multi-step workflows without synchronous human approval. It is not a chatbot. It is an always-on execution engine.

Independent academic research testing OpenClaw across 47 adversarial scenarios found a baseline native defense rate of just 17% against sandbox escape attacks. This is not a criticism of OpenClaw — it is a structural property of any LLM-based agent that relies on the model’s own safety training as its primary control. External enforcement is required.

The Gap Each Framework Leaves

SlowMist (standalone):

AI SAFE² (standalone):

Together:


Architecture Overview

┌─────────────────────────────────────────────────────────────────────────┐
│                      UNIFIED SECURITY ARCHITECTURE                      │
│                                                                         │
│   ORGANIZATIONAL LAYER (AI SAFE² Pillars 2, 4, 5)                       │
│   • Cross-deployment automation registry                                │
│   • Fleet-wide behavioral anomaly detection                             │
│   • Quarterly red-team exercises (SlowMist curriculum)                  │
│   • Annual threat model review                                          │
│                                                                         │
│   ┌─────────────────────────────────────────────────────────────────┐   │
│   │  GATEWAY LAYER — Control Gateway (AI SAFE² Pillar 4)            │   │
│   │  • Sits between OpenClaw ↔ LLM API                              │   │
│   │  • Real-time risk scoring, prompt injection blocking            │   │
│   │  • High-risk tool denial, immutable API-layer audit logs        │   │
│   └─────────────────────────────────────────────────────────────────┘   │
│                              │                                          │
│   ┌─────────────────────────────────────────────────────────────────┐   │
│   │  AGENT HOST LAYER — SlowMist Defense Matrix                     │   │
│   │                                                                 │   │
│   │  Pre-action: Red/Yellow Lines + Skill Audit                     │   │
│   │     └── Memory Vaccine (AI SAFE² Pillar 1) ◀── cognitive layer │   │
│   │                                                                 │   │
│   │  In-action: Permission Narrowing + Hash Baseline + Logs         │   │
│   │     └── Gateway enforcement continues here ◀────────────────── │   │
│   │                                                                 │   │
│   │  Post-action: Nightly 13-Metric Audit + Brain Backup            │   │
│   │     └── Vulnerability Scanner (AI SAFE² Pillar 2) ◀── here     │   │
│   └─────────────────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────────────────┘

Quick Start

Prerequisites

Step 1: Deploy the Memory Vaccine

Add openclaw_memory.md to OpenClaw’s memory bank. This is the cognitive-layer enforcement that complements SlowMist’s behavioral red/yellow lines.

# From the AI SAFE² examples/openclaw directory
cp openclaw_memory.md ~/.openclaw/workspace/
# Then in OpenClaw chat: "Please load and prioritize the security context in openclaw_memory.md"

Step 2: Run the Vulnerability Scanner (Baseline)

Establish your pre-deployment security baseline before activating any new skills.

wget https://raw.githubusercontent.com/CyberStrategyInstitute/ai-safe2-framework/main/examples/openclaw/scanner.py
python3 scanner.py --target ~/.openclaw

Review the 0–100 risk score and remediate CRITICAL findings before proceeding.

Step 3: Deploy the Control Gateway

# From the AI SAFE² examples/openclaw/gateway directory
# See gateway/README.md for full configuration options

Point OpenClaw’s LLM API calls through the gateway. The gateway handles all subsequent requests with real-time enforcement.

Step 4: Integrate with SlowMist Nightly Audit

Add the Scanner to SlowMist’s existing nightly cron:

# Append to your existing nightly-security-audit.sh, or add as a separate cron:
# 0 3 * * * python3 /path/to/scanner.py --target ~/.openclaw >> /tmp/openclaw/security-reports/scanner-$(date +%Y-%m-%d).txt

Route both SlowMist audit reports and Scanner output to your centralized log aggregator for cross-deployment visibility.

Step 5: Schedule Red-Team Exercises

Use the red-team-schedule.md in this directory to plan recurring adversarial exercises using SlowMist’s Validation Guide as your test curriculum.


The Five Pillars — Control Mapping Summary

AI SAFE² Pillar Primary SlowMist Controls AI SAFE² Tool Key Gap Closed
Sanitize & Isolate Red/yellow lines, skill installation audit, dedicated VM Memory Vaccine Persistent cognitive-layer enforcement; memory poisoning prevention
Audit & Inventory 13-metric nightly audit, hash baseline, skill baseline Vulnerability Scanner Secret exposure, network binding checks, 0–100 risk score
Fail-Safe & Recovery Human confirmation gates, brain backup, state/key separation Gateway circuit-breakers Automated pre-emptive halts; no 24h detection gap
Engage & Monitor Pre-flight checks, process/network audit, yellow-line counts Control Gateway Real-time API-layer enforcement; cross-agent anomaly correlation
Evolve & Educate Red-team validation guide, agent-native deployment, threat model docs Recurring drill schedule Codified exercise cadence; annual threat model refresh

For the complete control-by-control mapping, see safe2-for-slowmist-overlay.md.


Repository Contents

examples/slowmist-overlay/
├── README.md                        ← You are here
├── safe2-for-slowmist-overlay.md    ← One-pager: full pillar mapping + tool placement diagram
├── control-mapping-table.md         ← Complete cross-reference table
├── deployment-checklist.md          ← Step-by-step deployment order
├── threat-model.md                  ← Combined threat model
├── red-team-schedule.md             ← Recurring adversarial exercise schedule
└── resources.md                     ← Research papers, references, further reading

Contributing

This overlay is maintained by the Cyber Strategy Institute. Contributions that improve the mapping, add new SlowMist controls as they are released, or document deployment patterns from the community are welcome.

Please open an issue before submitting a PR for substantive changes to the architecture or pillar mappings.


References


License

This overlay documentation is licensed under CC-BY-SA 4.0, consistent with the AI SAFE² Framework methodology license. Code components follow MIT. See the root-level LICENSE files for details.


“If governance is not enforced at runtime, it is not governance. It is forensics.” — Cyber Strategy Institute


Repository navigation

Examples Index Framework Home Cross-Pillar Governance NEXUS Scanner MCP Profile

AI SAFE² v3.1 | Cyber Strategy Institute