CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
One is voluntary and American and tells you how to organize the work. One is binding and European and tells you what you owe, based on how risky your system is. Neither is hard. Being the person in the room who has actually read them is a career advantage that costs you one afternoon.
AI RMF 1.0 (NIST AI 100-1, published January 2023) is a voluntary framework from the US National Institute of Standards and Technology. It is a public-domain US Government work — you may adapt it freely with citation, which is why it is the safest backbone for a company policy or a capstone appendix. Its whole structure is four functions. Click each one.
A clean seven-minute walkthrough of the four functions from a well-known certification instructor — useful if you want the framework in someone else's words before you open the PDF. (Mike Chapple; approx. 7 minutes — verify the runtime on the page.)
NIST — AI Risk Management Framework (landing page) ↗
The hub: the framework, the companion Playbook, and the profiles NIST has published alongside it.
AI RMF 1.0 — the full PDF ↗
Around 40 pages of readable prose, no mathematics. Skim the four function sections; you will recognize most of it from your systems courses.
Regulation (EU) 2024/1689 is binding law, not guidance, and it is structured around a simple idea: obligations scale with risk. The same technology sits in a different tier depending on what you point it at. Click each band.
Tier diagram drawn for this course from the structure described in the Act. Placement of any specific product is a legal question answered from the text, not from a lecture diagram.
The AI Act — readable structure and article-by-article text ↗
A navigable presentation of the regulation. Useful for finding which article or annex actually governs a use case.
Key facts to keep straight: it is Regulation (EU) 2024/1689, published in the Official Journal in July 2024 and in force since 1 August 2024. Its obligations do not all start at once — different duties apply from different dates, and the schedule has been actively debated. Read the dates in the text you are citing, and say which version you read.
Three reasons this is still your problem. One: whether the Act reaches a given US-built product is a scoping question you answer from the regulation, not from a rule of thumb — and the answer is often yes for anything sold into or used from Europe. Two: big vendors comply globally because maintaining two products is expensive, so the requirements arrive in your procurement questionnaires either way. Three: the tier structure — obligations proportional to consequence — is simply a good way to think, and it is the structure US regulators, insurers, and enterprise buyers are converging on regardless of what any single legislature does next.
| NIST AI RMF 1.0 | EU AI Act — Reg (EU) 2024/1689 | |
|---|---|---|
| Who issues it | NIST, a US standards agency (Dept. of Commerce). Developed openly with industry input. | The European Parliament and Council. It is legislation. |
| Binding? | No — voluntary. Its force comes from contracts, procurement, insurers, and courts asking whether you followed a recognized practice. | Yes — binding regulation, with supervisory authorities and penalties. Consult the text for what applies and when. |
| What it asks of you | Organize the work: govern it, map the context, measure what you claim, manage what you find. It tells you the process, not the answer. | Classify the system by risk, then meet the duties attached to that tier — documentation, data and human-oversight requirements, transparency to the people affected. |
| Unit of analysis | Your organization and lifecycle. Who owns this, and how do risks get found and handled over time? | The system and its use case. What is this thing pointed at, and who bears the consequence? |
| Licence to reuse | Public domain (US Government work) — adapt with citation. | Official EU text — cite and link with attribution; not an open remix licence. |
| What it means for your capstone | Use its four functions as the headings of your Milestone II security section. It is free structure that a reviewer recognizes. | State the tier you believe your agent lands in and why, in one paragraph, citing the article or annex you relied on. Being wrong with reasoning beats being silent. |
The uncomfortable secret of AI governance is how little of it is new. Strip the vocabulary and it is the control environment your information-systems courses already taught, applied to a system that happens to hold a language model.
| What you built in this course | What the control is actually called | What governance adds |
|---|---|---|
| Tool inventory — every function the agent may call, with its scope (Week 5) | Asset inventory and least-privilege access control | A named owner per tool and a review date. An inventory nobody owns is a document, not a control. |
| Human gate before irreversible actions (Weeks 4–5) | Maker–checker / segregation of duties | Written criteria for what needs approval, and evidence that approvals actually happened. |
| Egress allow-list and output screening (page 01) | Data-loss prevention | A list of approved destinations that changes through a request, not a code commit. |
| Full step trace in the Week 2 evidence format | Audit trail | Retention, tamper-resistance, and someone whose job includes reading it. |
| Your eval harness and regression suite (Week 8) | Testing and control monitoring | A threshold that means “do not ship,” agreed before the number comes back. |
| Max-iteration guard, budget cap, kill switch (Week 4) | Operational limits and incident response | A documented plan for who gets paged, who can stop the agent, and what gets told to whom. |