🔒 Skill: Secure AI (v1.1.0)

Executive Summary

The secure-ai architect is the primary defender of the AI integration layer. In 2026, where AI agents have high levels of autonomy and access, the risk of Prompt Injection, Data Leakage, and Privilege Escalation is paramount. This skill focuses on building "Unbreakable" AI systems through multi-layered defense, structural isolation, and zero-trust orchestration.

📋 Table of Contents

1. Core Security Philosophies
2. The "Do Not" List (Anti-Patterns)
3. Prompt Injection Defense
4. Zero-Trust for AI Agents
5. Secure Server Action Patterns
6. Audit and Compliance Monitoring
7. Reference Library

🏗️ Core Security Philosophies

1. Isolation is Absolute: User data must never be treated as system instruction.
2. Least Privilege for Agents: Give agents only the tools they need for the current sub-task.
3. Human Verification of Destruction: Destructive actions require a human signature.
4. No Secrets in Client: All AI logic and keys reside in server-only environments.
5. Adversarial mindset: Assume the user (and the agent) will try to bypass your rules.

🚫 The "Do Not" List (Anti-Patterns)

🛡️ Prompt Injection Defense

We use a "Defense-in-Depth" strategy:

• Input Boundaries: --- USER DATA START ---.
• Guardian Models: Fast pre-scanners for malicious patterns.
• Content Filtering: Built-in safety settings on Gemini 3 Pro.

See References: Prompt Injection for blueprints.

🤖 Zero-Trust for AI Agents

• Non-Human Identity (NHI): Verifiable identities for every agent.
• WASM Sandboxing: Running generated code in isolated runtimes.
• HITL (Human-in-the-Loop): Mandatory sign-off for financial or data-altering events.

📖 Reference Library

Detailed deep-dives into AI Security:

• Prompt Injection Defense: Multi-layered isolation.
• Agentic Zero-Trust: Managing autonomous actors.
• Secure Server Actions: Bridging the frontend safely.
• Audit Protocols: Monitoring agent behavior.

Updated: January 22, 2026 - 20:50