ai-safe2-framework

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

Framework Home Research Index Cross-Pillar Governance AISM NEXUS Dashboard

Current framework context: AI SAFE² v3.1. This research note preserves its original publication date, evidence, and historical framework references. Use current v3.1 normative control and profile documents for implementation or conformance decisions.

Research Note: Runtime Isolation & The “Glass Box”

ID: RN-2025-006 | Related Control: [P1.T2.1], [P1.T2.1_ADV] | Status: Verified

🚨 The Threat Vector

Container Breakouts: Standard Docker containers share the host kernel. If an Agent has code-execution capabilities (e.g., Python REPL), a kernel exploit allows it to escape the container and compromise the host infrastructure.

🛡️ The AI SAFE² Solution

For Tier 3 (Agentic) systems, standard containerization is insufficient. We mandate User-Space Kernels.

1. gVisor / Firecracker Enforcement [P1.T2.1]

All agents capable of writing code must run in gVisor (runsc) or AWS Firecracker microVMs. This provides a distinct kernel boundary, mitigating escape vulnerabilities.

2. Ephemeral Runtimes

Agents should not have persistent filesystems. Containers must be destroyed and rebuilt after every task execution to prevent malware persistence.

📚 References


Research navigation

Previous research note Research Index Next research note
Framework Home Cross-Pillar Governance NEXUS Challenge Lab

AI SAFE² v3.1 | Cyber Strategy Institute