The Second NIST Cyber AI Profile Workshop highlighted a key issue in AI governance: the disconnect between enterprise-level policy and the engineers tasked with implementation.
The Challenge
NIST's Preliminary Draft of the Cyber AI Profile received over 1,400 comments during its public comment period. This unusually high number indicates a deeper problem: organizations struggle to operationalize AI governance for cybersecurity.
Your compliance team can draft AI governance policies, but if your security engineers can't translate "establish accountability for AI systems" into actionable technical controls, you're just adding to the documentation burden.
Workshop participants emphasized the need for both enterprise risk management resources and implementation-level guidance. Lacking either results in theoretical frameworks that are ignored or tactical controls that don't align with business risk.
The Environment and Constraints
A major difficulty is the lack of a consistent taxonomy for AI in cybersecurity.
When your security team mentions "AI-assisted threat detection," your compliance team might think "automated decision-making system," while your legal team considers "algorithmic processing under privacy regulations." Your engineering team refers to "machine learning model analyzing network traffic patterns." You're all describing the same system but can't agree on classification, governance, or risk measurement. This isn't just a communication issue; it's a structural gap that hinders coherent governance.
The workshop also raised the human-in-the-loop (HITL) question. If your AI system flags a security incident, who reviews it? Who can override it? What if the AI's recommendation conflicts with an analyst's judgment?
These are daily operational decisions your governance framework must address. Most AI governance policies jump from "maintain human oversight" to "document your decisions" without defining practical oversight.
NIST's Approach
The workshop's focus on adaptable guidelines reflects a practical reality: AI technologies evolve faster than standards cycles. Rigid requirements would be outdated before your first implementation is complete.
Instead, the Cyber AI Profile should provide enough structure for consistent risk assessment while allowing for your specific context. Focus on outcomes rather than prescriptive controls.
For instance, instead of mandating a specific type of monitoring for AI models, the profile should require that you can detect when an AI system's behavior deviates from its validated baseline. How you achieve this depends on your architecture, risk tolerance, and existing controls.
The taxonomy work is crucial for creating a common language. When NIST defines what constitutes an "AI system" in cybersecurity, your team can have consistent conversations with legal, compliance, and engineering. You can map these definitions to existing frameworks like NIST CSF v2.0 or ISO 27001.
On the HITL front, the framework should differentiate between types of human involvement:
- Human review of AI recommendations before action
- Human monitoring of AI actions with override capability
- Human audit of AI decisions after the fact
- Human retraining of models based on outcomes
Your governance structure should specify which level applies to which systems based on risk. An AI system prioritizing security alerts for human review requires less oversight than one that automatically blocks network traffic.
What's Still Missing
The workshop revealed gaps that NIST needs to address. Organizations seek clearer guidance on integrating AI governance into existing compliance frameworks. If you're managing controls for PCI DSS v4.0.1, SOC 2 Type II, and NIST 800-53 Rev 5, you don't want a separate AI governance program. You need to know which AI-specific controls map to your existing requirements.
The accountability question remains unresolved. When an AI system makes a security decision leading to an incident, who's responsible? The data scientist who trained the model? The engineer who deployed it? The security analyst who configured its parameters? The manager who approved its use?
You can't answer this with policy alone. You need technical controls that create an audit trail showing who made which decisions at each stage. This means logging model training data, deployment configurations, parameter changes, and human override decisions. Most organizations aren't doing this yet.
Takeaways for Your Team
Start with taxonomy. Before writing AI governance policies, define what counts as an AI system in your environment. Create categories based on risk and decision authority. A chatbot answering internal IT questions needs different governance than an AI system making access control decisions.
Map AI systems to your existing control framework. Don't create parallel governance. If you have a change management process for production systems, extend it to cover AI model updates. If you have incident response procedures, add scenarios for AI system failures.
Define HITL requirements based on impact. For each AI system, document:
- What decisions it makes autonomously
- What requires human review before action
- Who has authority to override it
- How quickly a human can intervene
Build the audit trail now. Even if you're still figuring out accountability, start logging AI system decisions, human overrides, and model changes. When regulators or auditors ask questions, you'll have data to work from.
Don't wait for the final Cyber AI Profile to act. The workshop's 1,400 comments show that organizations are already struggling with these issues. Start with the preliminary draft's structure, adapt it to your context, and iterate as NIST refines the framework.
Organizations that establish practical AI governance now will have an advantage when regulatory requirements tighten. Those waiting for perfect guidance will scramble to retrofit governance onto systems not designed for it.
