UK proposals cite two-step notification within 24 and 72

You must establish internal containment and evidence preservation immediately upon detecting an AI incident, then determine in parallel which statutory notification deadlines apply to your organisation and system. The applicable clock depends on your regulator, the system category and the nature of the harm — not on a single universal rule.

An analysis of 13 November 2026 of two-step notification deadlines for AI incidents argues that UK proposals tie qualifying incidents to a first notification within 24 hours and a fuller one within 72 hours, though only for certain regulated entities and subject to further threshold-setting in legislation. The UK AI Security Institute documented an incident in which monitoring flagged unusual data output, evaluations were stopped within about an hour and systems were isolated. In our assessment, the practical lesson is that triage, containment and evidence preservation must run in parallel with the legal determination of which notification duty applies, rather than sequentially — because waiting to establish all facts before stopping the system creates forensic loss and operational risk.

Which notification deadline applies to your incident?

The 24 and 72-hour structure proposed for UK cyber incidents applies only to entities falling within that scheme and remains subject to legislation. Northern Irish guidance on the EU AI Act cites immediate notification to the provider and market surveillance authority for serious incidents involving specific high-risk systems, with further reporting within 2 days for widespread breaches or critical infrastructure, 10 days for a death and 15 days in other cases — but only for systems falling under the Windsor Framework application. In our assessment, anyone treating these figures as universal AI rules draws an incorrect conclusion from context-specific guidance and thereby increases the risk of a flawed triage decision. Your lawyers should determine the applicable regime per AI system in advance, rather than working that out during the crisis.

What must you do in the first hour?

Stop the system involved immediately. Secure the evidence without changing or deleting anything before incident handlers have secured it. Record how and when awareness of the incident was established, including who made that determination, so the timing can be reconstructed and defended if a notification deadline is later questioned. Designate an authorised incident function able to direct stopping and evidence preservation without waiting for the investigation's outcome. Warn your privacy officer separately as soon as the involvement of personal data is plausible, because a separate 72-hour assessment under data protection law may apply independently of any AI-specific notification duty.

What evidence must you preserve?

For forensic reconstruction, secure at the point of containment:

• Model inputs and outputs — the data fed to the system and what it produced, to establish what the system actually did.
• System logs and audit trails — timestamped records of model execution, configuration changes and access.
• Agent transcripts and action sequences — if the system is an agent, the steps it took and the reasoning it recorded.
• Monitoring and alerting records — what triggered detection and when.
• Data lineage and provenance — which training data, fine-tuning or retrieval sources the system drew on.
• User interactions and prompts — the inputs that preceded the incident.

How should you structure parallel triage and notification assessment?

Run containment, evidence preservation, impact assessment and the legal notification assessment simultaneously from the first signal, so that none waits unnecessarily on the others. Document what you observe, not what you suspect — this distinction matters for both forensic accuracy and for defending your notification decision later. Because a single event may trigger several simultaneous notification duties under different regimes, record in the triage process which regulatory frameworks apply to each system involved and which deadlines each one carries. Avoid the assumption that every hallucination or model deviation is automatically notifiable; a notification duty depends on impact, system category, personal data involvement, critical services and ongoing harm.

Tooling can help you log, alert and preserve evidence automatically, but the determination of which notification duty applies and when awareness was genuinely established remains your own professional judgement. No system can make that legal assessment for you, and the cost of getting it wrong — either by notifying too late or by notifying when no duty applied — falls on you.