AI incident playbook: from preparation to lessons learned.
Guide G10 · Published
For organisations that use or build AI systems. What to do in advance, in the first hour, in the first day, during recovery and afterwards, and whom to tell in Ukraine.
Who it is for and when to use it
A guide for small and medium businesses, NGOs and civic organisations whose AI system, model, data or AI agent is behaving as if it were compromised, or really has been. NIST defines an AI incident as an event or series of events in which the development, use or malfunction of AI systems directly or indirectly contributes to harm: to people, critical infrastructure, human rights, property or the environment 2.
The guide covers four typical cases: a prompt injection that triggered actions; data leaking through a model; poisoned data or a poisoned model; and a hijacked AI agent or a compromised third-party model. It is not a plan for cyber incidents in general: ordinary breach-response measures apply here too, and this text adds what is specific to AI.
Prepare: what to do before an incident
- 01Name a person who can switch the system off, and decide in advance the conditions under which that happens. NIST advises having criteria for deactivation and a route to whoever decides 1 2.
- 02Turn on logging of system and user actions, and of changes to system prompts and model configuration: without it an incident cannot be investigated 3.
- 03Keep versions of models, data and configuration so that you can return to a known good state 3 7.
- 04Write an incident and recovery plan, test it and keep it current. For vendors, agree what happens during an incident in the contract 3 2.
- 05Set up contact with CERT-UA in advance, so that your first report is quicker 10. Prepare a contact list: vendor, lawyer, the responsible people in your organisation.
- 06Rehearse the plan. NIST advises regularly rehearsing incident response plans for third-party AI systems 2.
The template “AI risk self-assessment” can help you find gaps in your preparation.
The first hour
- 01Stop the harm, but do not destroy evidence. Switch off the agent or feature, or cut its access to other systems. NIST calls for mechanisms that can deactivate an AI system, and for people responsible for them 1. NAK recommends revoking the tokens and keys the agent used, or that could have entered the model's context.
- 02Preserve evidence before any clean-up: logs, prompts and responses, versions of prompts and configuration, the times of events. CERT-UA asks for logs and suspicious files 10.
- 03Tell whoever makes the decision, and start a timeline: when, who, what was done. NIST asks that tracking, response and recovery of incidents be documented 1.
- 04If there are signs that an attacker got into your network, isolate it to stop the spread: CERT-UA recommends this 10.
- 05Work out which of the four cases you are facing, and start with the matching steps below.
- Prompt injection that triggered actions. Find the content (a page, an email, a file) that carried the hidden instructions, and list the actions the tools carried out. The consequences depend on how much agency the model had 4. NAK recommends taking that source out of the model's reach.
- Data leaking through a model. Establish what data could have appeared in an output: personal, financial, commercial, credentials 5. A restriction in the system prompt can be bypassed, so do not assume it protected you 5.
- Poisoned data or model. Pause training, fine-tuning and updates of search knowledge bases from the suspect source, and check the origin and integrity of the data, for example with hashes 7 9. Poisoning can lie dormant until a specific trigger, so the absence of visible faults does not show that all is well 7.
- Hijacked agent or compromised third-party model. List the agent's tools and permissions and what it managed to do 8. Model repositories can be compromised and a supplier can be impersonated, so contact the real vendor through a channel you already know 6.
The first day
- 01Assess the scope from the logs: which data, users, systems and actions were affected. This is NAK's recommendation.
- 02Fix the cause, not only the effect: remove poisoned data, close the injection channel, cut the agent's permissions, change credentials 8 9.
- 03Tell the vendor as the contract provides. NIST advises deciding in advance who is responsible for incident response in third-party AI systems 2 3.
- 04Tell those affected: NIST advises communicating incidents to the relevant parties, including affected communities, and where applicable to legal and regulatory bodies 1 2.
- 05With a lawyer, decide whether any notifications are mandatory, and make them. See the next section.
- 06Keep the timeline going: you will need it for recovery and for the review.
Whom to tell in Ukraine
There is no single channel for AI incidents as such: NIST notes that formal channels for reporting AI incidents do not currently exist, and that there are only public databases that choose which cases to track 2. So use the ordinary cyber-security channels.
- Management, the system owner and a lawyer: inside the organisation, at once. This is NAK's recommendation.
- The model or AI-service vendor: through the contacts and terms of your contract 2 3.
- CERT-UA. Reports are accepted by email at incidents@cert.gov.ua and by phone on +38 (044) 281-88-25 (Mon–Thu 8:00–17:00, Fri 8:00–15:45; at weekends and on public holidays, around the clock: +38 (044) 281-88-01) 14. The affected organisation, or anyone who sees suspicious activity in a network, can report. CERT-UA is not a punitive body: its aim is to help organisations avoid incidents or recover from them, and it works with technical data (logs, suspicious files), not company documents 10.
- The Cyber Police, if there are signs of a crime. Its website has a form, “Report a cyber incident”, for malicious content (phishing, malware); other appeals go through cyberpolice.gov.ua/declare 11. A lawyer can help you decide whether your case is a crime.
- Personal data. If personal data may have been affected, bring in a lawyer immediately. The Ukrainian Parliament Commissioner for Human Rights receives notifications from data owners about processing that poses a special risk to the rights of data subjects 12. We did not find a ready procedure for reporting a data leak on those pages: whether you must notify the Commissioner and the people affected depends on your situation, and that is a question for a lawyer. The Commissioner's contacts, including a free hotline on 1678, are on its website 13.
- Public AI incident databases, if you wish. NIST mentions, for example, the AI Incident Database, AVID and the OECD AI Incident Monitor 2. NAK recommends not posting details there that expose victims or an unfixed vulnerability.
Recovery
- 01Return to a known good state of the model, data and configuration 3 7.
- 02Test the system before putting it back into service. The UK Code of Practice calls for appropriate testing and evaluation 3. NIST advises having procedures to fix problems and telling stakeholders the timelines 2.
- 03If data was involved, clean it and check it for anomalies before any retraining 9.
- 04Bring the system back with tighter permissions: the minimum of tools and rights, and human approval for high-impact actions 8 4.
- 05Monitor more closely for a period after recovery: analyse logs for anomalies and sudden changes in behaviour 3.
- 06Change every credential the compromised component could have reached. This is NAK's recommendation.
Lessons learned
- 01Hold a review with everyone who took part, to establish root causes and introduce measures that prevent a repeat 2.
- 02Check whether the channels to vendors and authorities worked, and update the incident plan and the notification procedure 2.
- 03Keep a record of errors, near misses and negative effects, not only of large incidents 2.
- 04Revisit your risk self-assessment: the gap that led to the incident belongs in your action plan.
Limits
This is a guide, not legal advice and not a service. NAK does not take incident reports, investigate incidents, recover data or analyse files. The steps are built on public guidance (NIST, OWASP, the UK Code of Practice, CISA), not on requirements of Ukrainian law: whether you have a duty to report depends on your situation, and that is a question for a lawyer. Contact details of public bodies can change, so check them on the official sites. This text was last checked on 24 Oct 2026; the sources used are listed below.
Sources
- 1NIST AI Resource Center. NIST AI RMF Core: Govern, Map, Measure, Manage · accessed 2026-10-24
- 2NIST. Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1) · accessed 2026-10-24
- 3UK Government (DSIT). Code of Practice for the Cyber Security of AI · accessed 2026-10-24
- 4OWASP GenAI Security Project. LLM01:2025 Prompt Injection · accessed 2026-10-24
- 5OWASP GenAI Security Project. LLM02:2025 Sensitive Information Disclosure · accessed 2026-10-24
- 6OWASP GenAI Security Project. LLM03:2025 Supply Chain · accessed 2026-10-24
- 7OWASP GenAI Security Project. LLM04:2025 Data and Model Poisoning · accessed 2026-10-24
- 8OWASP GenAI Security Project. LLM06:2025 Excessive Agency · accessed 2026-10-24
- 9CISA, 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
- 10Cabinet of Ministers of Ukraine. How CERT-UA responds to cyber incidents, from report to recovery · accessed 2026-10-24
- 11Cyber Police Department, National Police of Ukraine. Cyber Police: incident form and appeals · accessed 2026-10-24
- 12Ukrainian Parliament Commissioner for Human Rights. Notification of the Commissioner: procedure, forms and additional information · accessed 2026-10-24
- 13Ukrainian Parliament Commissioner for Human Rights. Ukrainian Parliament Commissioner for Human Rights: home page and contacts · accessed 2026-10-24
- 14CERT-UA. Contacts: report an incident · accessed 2026-10-24
Read next.
- Guide G8AI risk self-assessment: a template for organisations
- Guide G4AI security checklist for small organisations
- Guide G5Agentic AI: least privilege and human approval checklist
- Guide G1The first 24 hours after a deepfake of you or your company
- Guide G7Cyber preparedness minimum for NGOs and SMEs
- Guide G6Protecting your accounts and digital identity