SHIELD.md: A Security Standard for OpenClaw and AI Agents

The report proposes SHIELD.md v0, a context-loaded security policy for AI agents (e.g., OpenClaw) that integrates with MoltThreat to provide dynamic threat-driven decisions—log, require_approval, or block—across prompts, skill install/execute, tool calls, network egress, secrets read, and MCP, using explicit matching rules, confidence thresholds, and strict priority (block > require_approval > log). It outlines a minimal Decision block format, default behaviors, eligibility and enforcement constraints, and notes limitations (e.g., non-deterministic compliance, prompt injection risks, context limits), positioning SHIELD.md as early guardrails toward stronger agent security.