Determine where your organisation actually stands on cognitive security across four dimensions and five levels.
Score each one from 1 to 5, one at a time.
No document, record or number to point to? The score is capped at 2.
Set by your weakest dimension, with one priority to fix first.
Three ideas, one model
Three ideas that grew up together. The full explanation is in the UCSMM Field Guide.
Security that works quietly and protects the freedom to think and choose.
The human mind is the new perimeter: judgement, attention and choice. Human-factor anchor: ISO 45003.
Four dimensions, five levels, one honest assessment, a clear next step.
The core logic
Adopting AI turns into resilience only once you build the capability to protect it — and governance is what makes that turn happen.
Explore the model
Pick a tab, then click any item to read it. Short by design: the full detail is in the field guide.
Who owns each decision across the model lifecycle, under which document, and with what authority to stop. Not an ethics committee but a traceable accountability path: who authorised deployment, who reviews impact, and who can withdraw a model this week.
The ability to absorb an incident, keep operating, recover and learn from it. Protection assumes prevention is possible; resilience assumes failure and asks its cost and duration, with one question added for AI: how do we operate while our model is disabled or untrusted?
Moving from signature-based to behaviour-based detection: baselines for users, systems and models, with deviation raised as an alert. And what most organisations miss: monitoring the model itself, its output distribution, its accuracy on a reference sample, and the query patterns reaching it.
Changing controls at the speed of the threat, not the budget cycle. At its core an organisational learning capability: is every incident, lesson and exercise translated into a documented change to a control, a policy or a training programme? An organisation that learns faster than its adversary wins even when it is less well funded.
Fragmented, reactive practices, manual controls and inconsistent policy enforcement. AI systems, where they exist, run without dedicated security oversight.
Evidence required: None. The default level for every organisation not yet assessed.
Structured security policies, formally assigned responsibilities and basic monitoring. AI risk is recognised as a category in its own right, with early processes to manage it.
Evidence required: An approved policy · a named owner for AI risk · an initial model inventory.
Systematic integration of AI governance and security: policies protecting data and model pipelines, a unified security architecture, and AI risk assessment inside the development and deployment lifecycle.
Evidence required: A mandatory security review gate before any model deployment · a complete model register · training-data controls.
Security has become measurable: advanced analytics, behavioural monitoring, machine-learning-based detection, and performance indicators under continuous evaluation that anticipate weakness before it is exploited.
Evidence required: Time-based indicators (detect, contain, recover) · model drift monitoring · periodic board reporting.
A fully AI-augmented security function: autonomous learning algorithms, real-time threat intelligence and predictive analytics that detect and respond dynamically. The human analyst collaborates with the intelligent platform in an adaptive defence.
Evidence required: Documented automated response · proactive attack-scenario simulation · a learning cycle that changes controls without waiting for an incident.
Tools can produce an apparent level four on top of level-two governance, and that is more fragile than nothing, because it creates the illusion of safety.
Raise the weakest dimension before improving the strongest; enterprise resilience is set by its weakest dimension, not its average.
Claim no level without an auditable output behind it: a document, a record, or a measured indicator.
Inventory every AI system in use, including what business units adopted without security's knowledge. Assign an owner to each. Run the assessment and document each dimension's result. Draft a use and deployment policy.
The model register · a baseline assessment · an approved policy.
A pre-deployment security review gate, graduated by impact. Controls for training-data integrity and provenance. Model drift monitoring in production. An AI scenario in the response plan and one simulation exercise.
A documented gate procedure · a model monitoring dashboard · an exercise report.
Automate repetitive response paths. Enterprise-wide behavioural analytics. Proactive simulation of attack scenarios against models. A quarterly learning cycle that adjusts controls. Reassessment against the baseline.
Automated playbooks · an adversarial simulation report · a second, compared assessment.
Covers: Whether a control exists
Stops at: Not how intelligent it is; data integrity and model robustness are not assets in their own right.
Covers: Defined, stable assets
Stops at: A learning model is an asset whose behaviour changes after deployment.
Covers: Process consistency
Stops at: Designed for software processes: the procedure matures, the reasoning does not.
Covers: Who gets in
Stops at: Not whether a model's inference deserves trust; an authorised agent stays authorised while it is being misled.
Covers: An AI management system: policy, roles, risk and impact assessment
Stops at: It confirms the process exists, not how well the organisation detects, absorbs and adapts when a model misbehaves. UCSMM uses it as the backbone of the AI-governance dimension and measures the capability on top of it.
Covers: Psychosocial risk at work: the human under pressure
Stops at: Written for workplaces, not for systems that shape judgement; it protects wellbeing but not the freedom to decide when a model nudges. UCSMM carries it into the Cognitive Shield: the human factor is scored across all four dimensions.
Covers: Ethics and compliance
Stops at: More than adversarial behaviour, and rarely connected to security operations.
Each of them assumes a system behaves the way its documentation says. A learning system doesn't: its data changes, so its behaviour changes. UCSMM sits on top of what you already have and adds the missing piece.
Self-assessment
Score each statement from 1 to 5. Be honest: if you can't point to a document, a record or a number, the score stays at 2.
Before you begin
The assessment is a self-assessment against the model. Begin it and you are taking it on these terms:
About the model
My doctoral research, set out in the Applied Research Study (2026) · registered with the UAE Ministry of Economy and Tourism, certificate 3414-2026.
Elgendi, M. F. (2026). UCSMM: An applied research study for the AI-powered post-digital enterprise (Registration No. 3414-2026). Fawzooz. https://fawzooz.ai/ucsmm
