National Cybersecurity
Association of Ukraine

AI risk self-assessment: a template for organisations.

Guide G8 · Published

For organisations that use artificial intelligence (AI) tools or build their own AI features. Seven areas, each with a few questions and a description of what “good” looks like. It is a self-check, not a certification and not a legal assessment.

Who it is for and when to use it

The template is written for small and medium businesses, NGOs and civic organisations that use ready-made AI tools or build their own features or AI agents on top of them. Use it before launching a new tool, after significant changes (a new model, a new vendor, a new agent with access to your systems) and at regular intervals, for example every six months. The interval and what to do with the result are NAK's recommendations, not requirements of the sources.

How to work with it: answer each question “yes”, “partly” or “no” and note where the evidence lives (a policy, a setting, a log). The questions follow the OWASP list of risks for applications built on large language models 1, the functions of the NIST AI RMF 8 and the UK Code of Practice for the cyber security of AI 10. The “good” descriptions are NAK's recommendations built on those sources. Rank your “no” and “partly” answers by how much harm the gap could cause, and close the most dangerous first.

Do not enter secrets: passwords, API keys, tokens, or personal data. Not into this template, not into any other self-assessment tool, not into an AI chat. A “yes / partly / no” and a note on where the evidence is kept are enough. This page has no form, and nothing you answer is sent anywhere.

1. Data

  1. 01Do you know what data goes into AI tools: prompts, files, chat history, the knowledge bases the model searches? Good: there is a list of data types that must not go into external tools (secrets, personal, financial, health and confidential business data), and staff know it 3 10.
  2. 02Do you scrub sensitive data before it reaches the model? Good: sensitive content is masked or removed in advance, not left to an instruction in the system prompt, because such restrictions can be bypassed 3.
  3. 03Do you know the vendor's rules on retaining, deleting and using your data? Good: you know how long the vendor keeps data, whether it may use it for training, and whether you can opt out 3 4.
  4. 04Do you track where the data a model learns or searches from comes from, and how it changes? Good: sources are documented, changes are recorded, integrity is checked with checksums or hashes, and unverified sources do not enter the system without isolation 5 11.

2. Access and permissions

  1. 01Who can reach the AI tools, the data they read and their settings? Good: least privilege applies, meaning a person or process sees only the data needed for the work 3.
  2. 02Does the AI act as a specific user with that user's rights, rather than through a broad service account? Good: access to other systems is checked by those systems, not decided by the model 6.
  3. 03Where are secrets (keys, tokens) kept? Good: outside prompts and documents the model reads, because they count as sensitive data and a restriction in the prompt is not a reliable safeguard 3.
  4. 04Are request volumes and spending limited? Good: there are rate limits, quotas and usage monitoring, so abuse or a fault does not turn into runaway costs or an outage 7.

3. Vendors and third-party models

  1. 01Is there a current list of external models, APIs, plug-ins and datasets, with an owner for each? Good: the list is maintained, and OWASP recommends a dedicated bill of materials for it (SBOM/AIBOM) 4 10.
  2. 02Have you vetted the vendor? Good: you have reviewed its security posture and terms, including the privacy policy, and you would notice if they changed 4. For external components the Code calls for AI-specific risk assessment and due diligence 10.
  3. 03Do you check the source and integrity of models and files you download? Good: you take them only from sources you can verify and use signatures or hashes. A downloaded model can carry malicious code that runs when it is loaded 4 5.
  4. 04What happens if a failure or incident occurs at the vendor? Good: responsibilities are agreed in advance, contacts exist, and the response expectations are written into the contract 9 10.

4. Logging and monitoring

  1. 01Are system and user actions logged? Good: the logs let you reconstruct who asked what, when, and what the system did, well enough to support an investigation 10.
  2. 02Are changes to system prompts and model configuration recorded? Good: there is a change log showing who changed what and when 10.
  3. 03Does anyone look at the logs? Good: a named person or process looks for anomalies, leaks, signs of data poisoning and sudden changes in behaviour, and tool activity and resource use are tracked 6 7 10. NIST also asks for monitoring of the system once it is in production 8.

5. AI agents and tool use

  1. 01How many tools and permissions does the agent have? Good: only the ones it needs, narrow in purpose (for example a dedicated file-writing function, not a shell) and with the minimum permissions 6.
  2. 02Do high-impact actions need human approval? Good: yes, and a person approves the action before it is carried out 6 2.
  3. 03Do you treat the content the agent reads (web pages, emails, files) as untrusted? Good: it is separated and labelled, privileged actions are not triggered on its strength alone, and permissions are enforced in code, not in the prompt 2.
  4. 04Can you stop the agent quickly and limit the damage? Good: tool actions are logged, rates are limited, and there is a mechanism and a named owner able to switch the system off when its results contradict its intended use 6 8.

More on agents: the guide “Agentic AI: least privilege and human approval”.

6. Incident response

  1. 01Is there an incident plan that accounts for AI? Good: a written incident and recovery plan that has been created, tested and kept up to date 10.
  2. 02Is it decided who switches the system off, and under what conditions? Good: there are clear criteria for deactivation and a route to the person who decides 9 8.
  3. 03Can you restore a known good state? Good: versions of models, data and configuration are kept, and recovery planning covers attacks on an AI system 10 5.
  4. 04Do you know whom to tell outside the organisation? Good: the plan lists contacts for the vendor and public authorities, and contact with CERT-UA is set up in advance, since a first contact without preparation can take longer 12. For the response steps, see the guide “AI incident playbook”.
  5. 05Do you learn from incidents? Good: every case gets a review, and the plan is updated from the findings 9.

7. People and training

  1. 01Do staff know what may and may not go into AI tools? Good: there is a short acceptable-use policy, and training is tailored to each role 10.
  2. 02Does every AI system have an owner? Good: a named person is responsible for the system, its risks and the decision to switch it off 10 8.
  3. 03Do people know where to report it if the AI does something odd? Good: there is a simple channel that both staff and users can use 8.
  4. 04Are the people who build AI features trained in secure design and coding? Good: yes, and the AI security training is refreshed regularly 10.

Limits

This is a self-check, not a certification, not an audit and not a legal assessment. It does not prove that a system is secure or compliant with any law or standard, and there is no pass mark. NAK does not review your answers, run audits or investigate incidents. The list does not cover every risk: it leaves out, for example, copyright, bias and the rules on personal data. If the AI processes personal data, consult a lawyer. The UK Code of Practice is written for the United Kingdom and serves here as a reference, not as a requirement of Ukrainian law. The OWASP list is the 2025 edition: check whether a newer one has been published. This text was last checked on 24 Oct 2026; the sources used are listed below.

Sources

  1. 1OWASP GenAI Security Project. OWASP Top 10 for LLM Applications (2025) · accessed 2026-10-24
  2. 2OWASP GenAI Security Project. LLM01:2025 Prompt Injection · accessed 2026-10-24
  3. 3OWASP GenAI Security Project. LLM02:2025 Sensitive Information Disclosure · accessed 2026-10-24
  4. 4OWASP GenAI Security Project. LLM03:2025 Supply Chain · accessed 2026-10-24
  5. 5OWASP GenAI Security Project. LLM04:2025 Data and Model Poisoning · accessed 2026-10-24
  6. 6OWASP GenAI Security Project. LLM06:2025 Excessive Agency · accessed 2026-10-24
  7. 7OWASP GenAI Security Project. LLM10:2025 Unbounded Consumption · accessed 2026-10-24
  8. 8NIST AI Resource Center. NIST AI RMF Core: Govern, Map, Measure, Manage · accessed 2026-10-24
  9. 9NIST. Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1) · accessed 2026-10-24
  10. 10UK Government (DSIT). Code of Practice for the Cyber Security of AI · accessed 2026-10-24
  11. 11CISA, NSA, FBI and partners. AI Data Security: best practices for securing data used to train and operate AI systems (May 2025) · accessed 2026-10-24
  12. 12Cabinet of Ministers of Ukraine. How CERT-UA responds to cyber incidents, from report to recovery · accessed 2026-10-24