Direct sources. Clear status. Practical public information.
AI can help explain a security pattern, draft a test, or challenge an assumption—but a public AI service should not receive passwords, API keys, access tokens, private logs, customer data, proprietary code, or live configuration details. CISA’s public AI safety guidance says not to share sensitive or confidential information with AI systems. A safe review begins with a local redaction gate.
The four-column redaction matrix
| Field class | Examples to keep out | Safe substitute | Local proof after review |
|---|---|---|---|
| Secrets | Passwords, keys, tokens, cookies, recovery codes, private certificates | SECRET_A, fake key format, or synthetic credential |
Secret scanner and local secret-store check |
| Identity/private data | Names, email, phone, student/customer records, device IDs, exact IPs | USER_01, reserved example domains and addresses |
Authorized local mapping and privacy review |
| System detail | Private hostnames, topology, rules, vulnerabilities, production paths | Small synthetic architecture preserving only the design question | Owner-authorized local comparison |
| Business context | Private incidents, vendor terms, internal plans, unreleased code | Abstract scenario and fabricated sample | Human owner checks applicability and rights |
Build-first workflow
- Authorize. Name the system owner, scope, decision, allowed data, AI service, retention boundary, and human reviewer.
- Reduce locally. Remove secrets and identifiers before any external service; if the question cannot survive redaction, keep the review local.
- Synthesize. Build a minimal example that preserves the security property—validation, authorization, logging, error handling—without reproducing the live target.
- Ask defensively. Request threat assumptions, tests, safer patterns, or failure cases—not target-specific exploitation.
- Apply locally. Treat output as untrusted. Review dependencies, rights, privacy, and false-positive/false-negative risk.
- Repair and retest. Make the smallest controlled fix, rerun owner-authorized tests, and preserve evidence without secrets.
One safe practice task
Use an isolated sample application with fabricated users and reserved example values. Ask AI to propose tests for a missing authorization check. A human then implements the repair, runs the tests locally, confirms no real secret entered the prompt or output, and records the before/after result. Do not test a real third-party system without explicit authorization.
What changed in this edition
This edition adds a four-class redaction matrix and a six-step human-controlled review loop. It makes no claim that redaction eliminates all disclosure, retention, model, provider, legal, or cybersecurity risk.
Human stop point
Stop when authorization, data ownership, confidentiality, provider retention/training terms, export/IP rights, production scope, secret status, or testing permission is unclear. Keep the work local and involve the responsible security, privacy, legal, or system owner.
What Louisville Institute of Technology can contribute
LouisvilleIT can teach people to construct synthetic security cases, use AI without exposing live secrets, repair locally, retest, and preserve evidence. It does not authorize access, certify a system, replace professional review, or guarantee security.
Official sources
- CISA — Secure Our World: Using AI Tip Sheet
- NIST — AI Risk Management Framework
- NIST — Cybersecurity Framework 2.0
- OWASP — LLM07:2025 System Prompt Leakage
Sources observed November 1, 2026. Cybersecurity work is limited to systems and data the reader owns or is explicitly authorized to test.
Public-information notice: This Pulse is a dated, source-linked research summary for general educational use. It reports what the cited sources state or what the identified public record shows as of the date reviewed. Citation does not mean endorsement, partnership, certification, or independent confirmation of every third-party assertion. This is not individualized legal, medical, tax, accounting, investment, insurance, cybersecurity, licensing, regulatory, procurement, employment, real-estate, or other professional advice. Rules, facts, availability, and eligibility can change. Verify the controlling source, jurisdiction, effective date, and your own facts, and consult the responsible agency or a qualified professional before a consequential decision. External links are provided for attribution and may change. Use this information and any reader-controlled checklist at your own discretion and risk.