| 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.
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 |
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.
SlowMist (standalone):
AI SAFE² (standalone):
Together:
┌─────────────────────────────────────────────────────────────────────────┐
│ 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 │ │
│ └─────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────┘
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"
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.
# 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.
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.
Use the red-team-schedule.md in this directory to plan recurring adversarial exercises using SlowMist’s Validation Guide as your test curriculum.
| 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.
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
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.
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
| Examples Index | Framework Home | Cross-Pillar Governance | NEXUS | Scanner | MCP Profile |
AI SAFE² v3.1 | Cyber Strategy Institute