SOP - AI Safety and Data Discipline

SOP - AI Safety and Data Discipline (The “NetYeti” Way)

The Instructor's Perspective

We are using AI as a “force multiplier,” but a force multiplier only works if you don’t accidentally shoot your own foot. If we’re going to use LLMs—both Local (Primary) and Cloud (Alternate)—we need to maintain strict “signal discipline.”

Purpose and Scope

This SOP defines how we handle data when interacting with AI systems. It applies to all “intelligence missions” conducted within this journal and by the Cascade STEAM community.

1. The Golden Rule of AI Data

NEVER feed sensitive information into an AI that you wouldn’t want to see on a billboard.

• Local AI (P): While more secure, assume the data is still “warm” in memory.
• Cloud AI (A): Treat every prompt as public record.

2. Prohibited Data (The “No-Fly Zone”)

The following data points must NEVER be shared with any Cloud AI:

• Credentials: API keys, passwords, SSH private keys, or tokens.
• Personally Identifiable Information (PII): Real names, addresses, or phone numbers of students or peers without explicit consent.
• Internal Network Maps: Detailed IP addresses or firewall rules for private labs.
• Confidential Mentorship Notes: Private student progress reports or FERPA-protected data.

The Tactical Classification: Red-Yellow-Green

To keep it simple for the students at BTC, we use the “Traffic Light” system for data discipline.

Level	Data Type	Permitted Strategy
GREEN	Public docs, man pages, generic code, general “stupid questions.”	Cloud (A) / Local (P)
YELLOW	Internal IP schemes (sanitized), non-sensitive configs, draft documentation.	Local (P) ONLY
RED	Passwords, API Keys, SSH Private Keys, Student PII (FERPA), Network Maps.	OFF-LIMITS

The NetYeti's Teachable Moment

“A Manageable Mess is when you share a generic config and the AI gives you a bad answer. A Catastrophic Failure is when you share your .env file and your lab’s ‘keys to the kingdom’ end up in a training set for the next version of a model. Always check your clipboard before you hit Enter.”

3. Agentic Tool Safety (Aider & OpenClaw)

When using agentic tools that can “see” your filesystem, the risk of “Context Creep” increases.

• Aider: Use .aiderignore to explicitly block sensitive folders like .ssh, .git, or .env. Never run Aider in the root of a sensitive project without a “Safety Check.”
• OpenClaw: Ensure the allowlist is enabled and restricted to your specific NetYeti ID. Use the command-logger to audit what the agent is reading from your vault.
• Context Hygiene: Only provide the AI the files it needs to complete the mission. Don’t give it the whole “Armory” for a simple “recon mission.”

4. Signal Discipline (Sanitizing Prompts)

Before sending a prompt to the Alternate (A) layer (Cloud AI), perform the following:

• Redaction: Replace “192.168.1.50” with <INTERNAL_IP>.
• Anonymization: Use generic roles like “User A” or “Student 1.”
• Generalization: Instead of “I have a Cisco Catalyst 2960 with a VLAN issue,” use “I have a Layer 2 switch with a VLAN issue.”

4. Verification & Validation (AAR)

AI is a “hallucination machine.” Every output must be treated as a “draft” until:

• Code: It is run in a sandbox or isolated LXC.
• Knowledge: It is cross-referenced with official documentation (RFCs, man pages).
• Strategy: It is reviewed for common sense and “NetYeti” philosophy.

5. Local LLM (Primary) Benefits

We prioritize the Primary (P) layer (Local Ollama) for:

• Privacy: Data never leaves our hardware.
• Control: We own the weights and the inference engine.
• Resilience: The “brain” works even when the internet doesn’t.

Compliance and Consequences

Failure to follow signal discipline results in a “manageable mess” becoming a “catastrophic mess.”

• Correction: Immediate redaction of the note and a “lessons learned” entry in the journal.
• Prevention: Update the .env and .gitignore to ensure secrets are never even near the AI’s “sight.”

Common Pitfalls (The “Manageable Mess”)

• The “Context Creep”: Providing just enough redacted info that the AI can still figure out the context (e.g., “A specific tech college in Bellingham”).
• Local AI Overconfidence: Thinking “Local AI is 100% safe.” It’s safe from the cloud, but if your machine is compromised or if you log the prompts to a public Git repo, you’ve still lost signal discipline.
• Copy-Paste Fatigue: Accidentally pasting a .env file because it was still in your clipboard. Always Look Before You Enter.

Knowledge Check

• What is the “Golden Rule” of AI data?
• Why do we prioritize the Primary (P) layer for sensitive “intelligence missions”?
• What should you do before sending a technical question to a Cloud AI (A) layer?

---

Related: AI Command Center, Secret Management SOP