As NIST develops the Cybersecurity Framework Profile for Artificial Intelligence (Cyber AI Profile), risk management and cybersecurity teams are figuring out what this means for their programs. The January workshop drew over 1,400 comments, highlighting key concerns: managing AI risks without overhauling your framework, assigning accountability when AI systems make decisions, and evaluating if your current testing approach works for AI.
Here's what you need to know.
Can I Use My Existing Framework for AI?
You don't need to start over. The Cyber AI Profile is designed to integrate with NIST Cybersecurity Framework (CSF) 2.0, allowing you to adapt your current risk management approach.
The Profile highlights three Focus Areas where AI introduces specific considerations: securing AI systems, using AI for cybersecurity operations, and managing Automated Individual Decision-Making and Profiling risks. Your existing CSF implementation already covers most control categories. The change lies in how you apply them.
For instance, if you're using AI to analyze security logs, your asset management processes (CSF Identify function) should treat training data as a critical asset. Your incident response procedures (CSF Respond function) should address scenarios where an AI-driven detection system produces false negatives. You're extending your current functions to cover AI-specific scenarios.
If you're resource-constrained, start by mapping where you're already using AI, including "shadow AI" deployed without formal approval, and apply your existing controls there first.
Who's Accountable for AI-Related Security Incidents?
Accountability is still being defined. Some organizations are forming multidisciplinary AI governance bodies or creating chief AI officer roles, but there's no consensus yet on accountability for AI cybersecurity outcomes.
Document decision rights explicitly. When your AI-driven access control system denies a legitimate user, who investigates? When your AI tool misses an intrusion, who's responsible for the gap analysis? These are operational realities you need to plan for.
Human-in-the-loop processes remain the standard for AI accountability. This means defining escalation paths and human oversight for high-stakes outcomes. If your AI system flags a transaction as fraudulent and blocks it, someone needs to own the review process and criteria for overriding the AI's decision.
Your Statement of Applicability or risk register should address Automated Individual Decision-Making and Profiling authority. Which AI systems operate autonomously? Which require human approval? What are the thresholds for escalation? Document this before an incident forces you to explain it to regulators or auditors.
How Do I Test AI Systems Effectively?
Testing and evaluation are common challenges. Traditional penetration testing and vulnerability scanning don't capture AI-specific risks like model drift, adversarial inputs, or biased training data.
Start with measurable metrics. Track false positive and false negative rates for AI systems used in anomaly detection. Significant changes in these rates indicate drift that requires investigation.
For third-party AI tools, ask vendors about their testing methodologies. Do they provide certifications or audit reports? Can they demonstrate model performance validation? AI Bills of Materials (AIBOM) are emerging for supply chain transparency. If your vendor can't tell you what data their model was trained on or how they validate its outputs, that's a red flag.
Consider testing for adversarial resilience. Can inputs be manipulated to produce incorrect outputs? For example, if you're using AI for document classification, test whether subtle changes to document formatting cause misclassification. This requires thinking like an attacker, not deep ML expertise.
Should I Wait for the Final Profile Before Making Changes?
No. The Initial Public Draft is coming, but you're already managing AI risks. The preliminary draft and workshop feedback provide enough direction to start.
Focus on visibility first. Address "shadow AI" and lack of visibility into AI system training. If you don't know which teams are using AI tools or which business processes depend on AI decisions, you can't manage the risk.
Implement basic inventory controls: require teams to register AI tools, document data access, and identify business processes dependent on AI outputs. This isn't about blocking innovation; it's about knowing what you're managing.
Then apply your existing risk assessment process. Evaluate each AI system using the same criteria you'd use for any critical technology: What data does it access? What decisions does it make? What happens if it fails or produces incorrect outputs? What's the recovery process?
What's the Biggest Mistake in AI Governance?
Treating AI as purely a technology problem. AI governance requires multidisciplinary approaches, but many organizations still silo AI oversight within IT or data science teams.
Your legal team needs to be involved because AI decisions create liability exposure. Your HR team needs to be involved if AI systems make employment-related decisions. Your procurement team needs to be involved because AI vendor contracts require specific language around data rights, model transparency, and accountability.
Start with policy, then implement technical controls that enforce it. The Cyber AI Profile will give you a taxonomy for communicating about AI risks across disciplines, but you need the governance structure to act on that communication.
Where Can I Get More Information?
NIST's Cyber AI Profile project page maintains the latest roadmap and draft materials. The workshop recording and slides are available for detailed discussions.
If you're implementing ISO/IEC 42001 or working with the NIST AI Risk Management Framework, these efforts should inform your Cyber AI Profile implementation. These frameworks complement each other.
Join NIST's Community of Interest for notifications about working sessions and the Initial Public Draft release. The 1,400+ comments they received show this is a collaborative process, and your input will shape the final guidance.
The technology is moving fast, but your risk management fundamentals haven't changed. Identify your assets, assess your risks, implement controls, and monitor for effectiveness. The Cyber AI Profile gives you a structured way to apply those fundamentals to AI. Start now with what you know, and refine as the guidance evolves.
