
AI in the incident response plan: who acts in an emergency?
AI-related incidents need clear responsibilities. How organisations extend existing security procedures to cover reporting, intervention and recovery.
Philipp Kappe
Founder of IMScollect
Contents
A deceptively genuine email, a fake phone call or a suspicious action by an AI application: in an emergency, a general reference to “IT” does not help. It must be clear who receives the report, limits the damage and decides on the next steps. The AI security study published in October by the Federal Office for Information Security (BSI) and the TÜV Association (TÜV-Verband) provides a concrete reason to review existing procedures.
What the study actually says about incident processes
According to the AI Security Study 2026, 11 per cent of companies with existing or planned use of AI have specific processes for AI-related security incidents. A further 43 per cent rely on general IT security processes, so far without specific AI rules. 25 per cent say they have no arrangements for this yet. The study was published on 7 October; trade press coverage from 28 October picks up the results again.
The subject of the question is important: at this point, the study report, page 47, deals with the response to cyber attacks or fraud attempts in which perpetrators use AI. The figure is not proof that all other companies are unprotected or that they lack any emergency procedure at all. Nor can it be directly inferred from this how well companies handle disruptions to their own AI applications.
The entire forsa survey covers 505 companies with 20 or more employees in Germany. Those surveyed were people responsible for AI, IT security or IT management, or members of executive management, between 22 June and 31 July 2026. The percentages are self-reported, not results of technical security tests.
Distinguishing two kinds of AI involvement
For your own organisation, it is worth going through two perspectives separately. First: attackers use AI, for example for credible phishing messages or fake voices. The emergency procedure does not have to be completely reinvented for this. But call-back procedures, payment approvals and internal reports should work even when language and manner seem convincing.
Second: your own AI application is involved in the incident, for example because a connected service performs unusual actions or confidential content appears in an output. This is a different scenario with different options for intervention. Who is allowed to block access, disconnect a connection or temporarily deactivate the application? Which operational processes depend on it? These questions should be clarified in advance with those responsible on the technical and business side.
For AI agents, OWASP names, among other things, limited tool permissions, human control of high-risk actions and traceable logging. These are technical protection principles. An operational response procedure adds responsibilities and decisions to them; it does not replace them.
Adding five decisions to the existing procedure
The following overview is an organisational working aid, not a new statutory or normative list of obligations. First test it against an application that is actually in use or a specific fraud scenario.
| Decision | Clarify in advance | Record in the case |
|---|---|---|
| Who takes the report? | Reporting channel, reachable person and deputy | Observation, time and affected application or business transaction |
| Who limits the damage? | Technically authorised team and business approval for interventions | Stopped transaction, blocked access or other immediate action, including its consequences |
| Who secures evidence? | Suitable access to logs, messages and configuration states | Locations, responsible person and secured state |
| Who assesses further consequences? | Involvement of information security, data protection, the business unit and legal advice where applicable | Affected data, processes, external bodies and decisions taken |
| Who releases operations again? | Criteria, reviewers and person with decision-making authority | Test result, remaining restrictions and release |
An example: a supposed call from the managing director demands an unusual payment. A sensible first step may be to suspend the payment and verify the instruction via an already known, independent contact channel. Nobody first has to prove that the voice was AI-generated. The further course of action depends on the facts and the existing security process.
Even with a suspicious AI application, simply switching it off is not always the right immediate action. Interventions must limit damage while taking into account important evidence and operational dependencies. Such decisions belong in the hands of the responsible people, not in a general chatbot instruction.
Rehearsing the procedure instead of creating another folder
A short tabletop exercise often makes gaps easier to understand than an additional policy text. Use only fictitious or suitable test data, no real secrets. A possible scenario: an external service provider reports that a supposed internal instruction came from an unknown address and requests unusual data.
Ask the participants to explain what exactly they would do. Where is the reporting channel? Who can stop the affected transaction? Which person can provide logs? Who talks to the service provider? Document unresolved responsibilities and missing access, not just participation in the exercise.
Derive a few actionable improvements from this, with responsible persons and dates. An updated contact list, a reliable call-back route or tested access to logs can be more helpful than another abstract rule. Then use the same scenario to check whether the change actually works.
Conclusion
AI-related incidents need people who are able to act, clear options for intervention and a traceable procedure. What matters is which existing security processes help and which additions the specific scenario needs. A low rate of specific AI processes implies neither a blanket lack of protection nor the need for a completely new parallel system.
Risk and action planning is a sensible place to link identified gaps with responsibilities and improvements. IMScollect supports the organisational documentation of risks, actions and responsibilities. It does not replace technical detection, access blocking or switching off external AI applications. A demo with your own procedure can show how organisational follow-up can be mapped for your specific case.
Frequently asked questions
Does every company need its own AI incident response plan?▾
Not necessarily as a separate document. What matters is whether the existing incident processes sufficiently cover the relevant AI-related scenarios, options for intervention and responsibilities; the specific additions depend on the use and the risk.
Does the study's 11 per cent figure mean that all other companies are unprotected?▾
No. The figure refers to specific AI incident processes at companies with existing or planned use of AI. 43 per cent report general IT security processes without specific AI rules; no blanket lack of protection can be inferred from this.
Who should decide in an AI-related security incident?▾
Responsible persons named in advance, with the necessary business and technical authority, and a deputy. Depending on the facts, information security, the affected business unit, data protection and, where applicable, legal advice must be involved.
How can an AI-related incident procedure be tested?▾
Run through a specific scenario with the responsible people involved, without using real secrets or production data. Record unresolved responsibilities, missing access and necessary improvements, and test the procedure again after the changes.



