Zero Trust

**Judgement:** A strong, security-focused behavioral protocol that is well-suited as a default safety layer for agents performing any risky or externally-visible actions. **What it does well** - Enforces a strict zero-trust posture: all external inputs and actions are treated as potentially malicious until verified. - Provides a clear decision flow (STOP → THINK → VERIFY → ASK → ACT → LOG) that can be consistently applied. - Distinguishes between high-risk actions that require explicit human approval (e.g., unknown URLs, emails, financial actions) and low-risk, local actions that can proceed freely. - Gives concrete, practical rules for package installation safety and credential handling (no logging/printing, proper storage and permissions, no external sharing). **Risks / Limitations** - Human-in-the-loop requirement for many actions can significantly slow or block fully automated workflows. - It is a policy/protocol; actual safety depends on correct implementation and adherence by the agent runtime (it does not itself enforce technical sandboxing or isolation). - Some edge cases may still be ambiguous (e.g., what counts as a "known" service or "trusted" source) and require additional, environment-specific policy. **Recommended scenarios** - As a baseline security skill enabled by default for any general-purpose agent that can touch the network, handle credentials, or perform irreversible operations. - For organizational or production deployments where preventing accidental data leakage or risky external actions is more important than maximum automation speed. - As a complementary control alongside more technical defenses (sandboxing, rate limits, allowlists) to improve overall operational security of agents.