Adopt AI in cybersecurity in seven steps: set ownership and policy, pick one narrow use case, check data and access, secure the AI system itself, pilot beside the current process, measure against a baseline, then expand. NIST's draft Cyber AI Profile organizes this around three focus areas: secure, defend and thwart.
What does NIST's draft profile say?
The NIST preliminary draft Cyber AI Profile, IR 8596, is organized by the outcomes of Cybersecurity Framework 2.0. It addresses three focus areas: securing AI system components, conducting AI enabled cyber defense, and thwarting AI enabled cyber attacks. The draft says it can help organizations strategically adopt AI while addressing and prioritizing the cybersecurity risks that AI introduces. It is preliminary, with a comment period that has closed, so use it as direction and watch for later versions.
CSF 2.0 itself has six functions: Govern, Identify, Protect, Detect, Respond and Recover. Govern sits first because policy, roles and risk appetite decide everything after it, and that is where AI adoption should begin too.
Step by step
Step 1. Assign ownership. Name who is accountable for AI use in security, who approves new use cases, and who reviews incidents involving AI. Write a short policy on what data may go to which tools.
Step 2. Choose one use case. Pick a narrow task with a measurable result and low harm if wrong, such as alert grouping or drafting ticket text. Avoid starting with automated response.
Step 3. Check data and access. Decide what the system will read. If it reads logs, tickets or code, confirm that sensitive data is handled according to policy, and that vendors do not retain it in ways you did not agree to. Give the system its own identity with narrow permissions.
Step 4. Secure the AI system. This is the Secure focus area. Treat the model, its prompts, its tool connections and its training or reference data as assets in your inventory. Control who can change them, log what the system does, and test how it behaves when fed hostile input.
Step 5. Pilot beside the current process. Run the AI output next to the human process for several weeks. Compare verdicts, time spent and errors, and let analysts mark where it helped or misled.
Step 6. Measure against a baseline. Record the numbers before the pilot: time to triage, time to close, false positive rate, analyst hours. Decide up front what improvement justifies keeping the tool.
Step 7. Expand deliberately. Add the next use case only when the first has met its measure. Keep approval gates on irreversible actions, and review the whole setup on a schedule.
Why does the defensive side need the secure side?
Because an AI tool in your security stack is also a system an attacker may target. It reads untrusted text, it may hold credentials, and it can act. The Secure focus area exists for that reason. Adopting a model without securing it adds an attack path to the very program meant to reduce risk.
A constructed example
A security team of eight picks vulnerability triage as its first use case. Step 1 names the head of security operations as owner and records that source code may not leave the company's approved environment. Step 3 gives the triage tool read access to the dependency inventory and ticket system, and no write access to production. In the pilot, the tool's priority labels match senior analyst judgement on most tickets but misses one where a flaw is exposed through an internal gateway. The team adds exposure information as an input before expanding.
How to check or apply this
- Write the owner and policy in a page, not a slide.
- Pick one use case and a measure you can collect today.
- Map the data the tool will read and the permissions it needs, then cut both to the minimum.
- Add the AI system to your asset inventory and your change control.
- Pilot beside the human process and record disagreements.
- Decide the keep, change or stop threshold before the pilot ends.
- Review quarterly, and read new NIST versions as they appear.
Quick check
Where should an AI adoption plan begin?
With governance: an owner, a policy and a use case with a measure. Tools come after those decisions, not before.
Is the Cyber AI Profile a requirement?
No. It is a NIST community profile, currently a preliminary draft, offered as voluntary guidance.
Sources
- NIST IR 8596 preliminary draft, Cybersecurity Framework Profile for Artificial Intelligence
- NIST Cybersecurity Framework 2.0, CSWP 29
- NIST AI Risk Management Framework
Last reviewed: 18 October 2026.
