AI system incidents differ fundamentally from traditional software failures. Article 73(6) mandates evidence preservation before any system alteration, requiring dedicated response procedures that address ambiguous root causes, diffuse impact, and perishable forensic evidence.
Traditional incident response frameworks assume clear failure modes, but AI systems fail in subtler ways: systematic bias, accuracy degradation for specific subgroups, or outputs never observed during testing. This guide establishes an AI-specific incident response framework covering six incident categories mapped to the threat model: model performance degradation, fairness drift, data poisoning, adversarial exploitation, privacy breach, and human oversight failure. Each category requires defined detection mechanisms, severity classification, and pre-approved containment actions. Article 73(6) prohibits altering the AI system before informing competent authorities, making evidence preservation legally mandated and time-critical. The framework covers all response phases from detection through post-incident review, with clear role assignments including Incident Commander, Technical Lead, Communications Lead, Legal Liaison, and Evidence Custodian. Testing through annual tabletop exercises and biannual live simulations validates the plan before a real incident occurs. For organisations without dedicated tooling, procedural alternatives using manual on-call rotas and structured review templates provide a baseline capability, though automated alerting platforms are strongly advisable for time-sensitive systems.
AI system incidents differ fundamentally from traditional software failures, requiring a dedicated incident response framework that addresses characteristics unique to machine learning systems.
AI system incidents differ fundamentally from traditional software failures, requiring a dedicated incident response framework that addresses characteristics unique to machine learning systems. Traditional incident response assumes clear failure modes: a service goes down, a database becomes corrupted, a security breach is detected. AI incidents present subtler patterns where the system remains operational but produces systematically biased predictions, degrades in accuracy for specific subgroups, or generates outputs never observed during testing.
The root cause of an AI incident may be ambiguous. An unexpected output could stem from a model deficiency, a data pipeline error, a feature engineering bug, an input, a deployment configuration mistake, or a legitimate but unusual input the model was not trained to handle. Diagnosis demands a different skillset from traditional software debugging, combining ML engineering expertise with data analysis and domain knowledge. The incident response plan must therefore be specific to the and must account for these AI-specific characteristics that conventional frameworks do not cover.
The incident response plan should define AI-specific incident categories that map directly to the threat model established during the Cybersecurity for AI Systems risk assessment.
The incident response plan should define AI-specific incident categories that map directly to the threat model established during the Cybersecurity for AI Systems risk assessment. Each category requires a defined detection mechanism, severity classification, and response procedure.
Detection relies on the monitoring infrastructure operating continuously across all deployed AI systems.
Detection relies on the monitoring infrastructure operating continuously across all deployed AI systems. Automated alerts from the performance monitoring system, reports from human operators, deployer complaints, affected person feedback, and external disclosures such as vulnerability reports or media coverage all feed into the detection process.
The monitoring layer should generate category-specific alerts: a performance metric crossing a threshold triggers a performance degradation alert; a fairness metric crossing a threshold triggers a fairness drift alert; an anomalous pattern in inference logs triggers an adversarial exploitation alert. The engineering team integrates this alerting with the incident management platform so that alerts reach the on-call team within minutes.
Evidence preservation is time-critical and legally mandated.
Evidence preservation is time-critical and legally mandated. Article 73(6) explicitly prohibits altering the AI system in a way that could affect subsequent evaluation of the causes before informing the competent authorities. The evidence preservation procedure must therefore be executed before any model rollback, configuration change, or system modification, unless the system is actively causing harm.
A model's state, configuration, and inference logs constitute the forensic evidence needed to understand what happened. Upon incident detection, the response team must capture the model version and configuration at the time of the incident, the specific input data and feature values for affected decisions, the system's output for those decisions, the complete logging trail for the affected time period, and the monitoring metrics showing the anomaly.
Upon incident detection, the engineering team triggers an automated snapshot script that captures all forensic evidence before remediation begins.
Upon incident detection, the engineering team triggers an automated snapshot script that captures all forensic evidence before remediation begins. The script records the currently deployed model version from the model registry, the current configuration from the config management system, and the inference logs for the incident period from the logging infrastructure.
It also captures the monitoring metrics for the incident period from the monitoring platform and the current data pipeline state from the orchestration tool. These snapshots should be written to immutable storage such as S3 Object Lock or Azure Immutable Blob Storage immediately, preserving evidence before any remediation actions alter the system state.
Containment for AI incidents often differs from traditional containment approaches.
Containment for AI incidents often differs from traditional containment approaches. Where a traditional incident might be contained by isolating a compromised server, an AI incident may require rolling back the model to the last known good version, suspending automated decision-making and routing all cases to human review, temporarily shutting down the inference service, or restricting the system to a subset of its intended purpose while the affected functionality is investigated.
The incident response plan should define containment actions for each incident category, with pre-approved authority for the on-call team to execute containment without waiting for management approval. Time is critical: every minute the system operates with degraded fairness or compromised integrity is a minute during which affected persons may be harmed. Where the system is actively causing harm, the break-glass procedure is activated simultaneously with evidence capture.
Before the system returns to full production, the corrected version must pass the complete validation suite, including performance, fairness, and robustness gates.
Before the system returns to full production, the corrected version must pass the complete validation suite, including performance, fairness, and robustness gates. Recovery should follow the staged deployment process of staging validation, canary deployment, and full rollout to limit exposure if the fix introduces new issues.
The Technical SME intensifies post-recovery monitoring for a defined period, typically two to four weeks, to confirm the issue does not recur. This extended observation window accounts for the possibility that AI failures may manifest slowly, with degradation only becoming apparent after the model processes a sufficient volume of diverse inputs.
The incident response process must integrate seamlessly with the serious incident reporting obligations under Article 73.
The incident response process must integrate seamlessly with the serious incident reporting obligations under Article 73. The triage phase should include an explicit assessment of whether the incident meets the Article 3(49) definition of a serious incident. If it does, or if there is a reasonable likelihood, the reporting timeline begins immediately.
The incident response plan must define clear role assignments with named individuals and alternates for out-of-hours coverage.
The incident response plan must define clear role assignments with named individuals and alternates for out-of-hours coverage. The Incident Commander, typically a senior Technical SME, holds authority over the entire response. A Technical Lead handles diagnosis and remediation. A Communications Lead manages deployer and authority notifications. A Legal Liaison, the Legal and Regulatory Advisor, assesses regulatory reporting obligations. An Evidence Custodian preserves and secures all incident evidence.
These roles should be pre-assigned so the team can mobilise immediately when an incident is declared. Named alternates ensure coverage during absences, weekends, and holidays. Role clarity prevents confusion during the high-pressure early minutes of incident response. The triage outcome determines which roles are activated: a minor performance degradation may require only the Incident Commander and Technical Lead, while a serious incident triggering Article 73 reporting requires the full team including the Legal Liaison.
The Technical SME tests the incident response plan through tabletop exercises at least annually and through live simulation exercises at least biannually.
The Technical SME tests the incident response plan through tabletop exercises at least annually and through live simulation exercises at least biannually. Tabletop exercises walk through hypothetical scenarios with the response team, testing decision-making and coordination without impacting production systems.
Live simulations inject controlled anomalies into a staging environment and exercise the full response process from detection through recovery. These simulations validate that alerting reaches the right people, that evidence preservation captures all required artefacts, that containment actions execute correctly, and that the recovery process restores the system to a validated state. Exercise results, including identified weaknesses and improvement actions, are documented by the Technical SME as Module 9 evidence.
Every incident should produce a post-incident review document following a blameless format that focuses on systemic causes and improvements rather than individual fault.
Every incident should produce a post-incident review document following a blameless format that focuses on systemic causes and improvements rather than individual fault. The review should cover the incident timeline, the detection mechanism and detection latency, the containment actions and their effectiveness, the root cause analysis, the impact assessment including how many affected persons were affected and how, and the corrective actions to prevent recurrence.
The review should assess whether the monitoring infrastructure detected the incident as quickly as it should have, whether the response followed the documented procedure, and whether any changes to the system, the monitoring, or the incident response plan are warranted. Findings feed back into the risk register for new risks identified, the AI system documentation for necessary updates, and the cybersecurity testing programme for new test cases. The report is retained as Module 6 evidence.
Incident response is fundamentally a process, not a tool.
Incident response is fundamentally a process, not a tool. Platforms such as PagerDuty and Jira Service Management automate the workflow, but the process can be executed manually as a procedural alternative.
A manual approach requires an incident response plan document covering detection criteria, triage process, containment actions, communication plan, evidence preservation steps, and post-incident review procedure. The Technical Owner maintains an on-call rota in a shared spreadsheet or calendar. An incident log in spreadsheet or document form records each incident with its timeline, actions taken, evidence collected, and resolution. A post-incident review template includes mandatory sections for timeline, root cause, impact assessment, and corrective actions.
Immediately. Evidence snapshots should be written to immutable storage before any remediation actions, as Article 73(6) prohibits system alterations that could affect evaluation of causes before authorities are informed.
Containment stops ongoing harm through actions like model rollback or routing cases to human review. Eradication addresses the root cause by correcting corrupted data, fixing pipeline bugs, retraining drifted models, or patching vulnerabilities.
Typically two to four weeks of intensified monitoring to confirm the issue does not recur, accounting for the possibility that AI failures may manifest slowly across diverse inputs.
AI incidents have ambiguous root causes, diffuse impact across many decisions, and perishable evidence that traditional frameworks do not address.
Six categories: model performance degradation, fairness drift, data poisoning, adversarial exploitation, privacy breach, and human oversight failure.
Article 73(6) prohibits altering the AI system in a way that could affect evaluation of causes before informing competent authorities.
Through tabletop exercises at least annually and live simulation exercises at least biannually, with results documented as Module 9 evidence.
Incident Commander, Technical Lead, Communications Lead, Legal Liaison, and Evidence Custodian, with named alternates for out-of-hours coverage.
Yes, through manual processes with documented plans, on-call rotas, and structured review templates, though automated alerting is strongly advisable.
Impact is often diffuse rather than concentrated at a single point of failure. An AI system may produce subtly incorrect outputs across many decisions before the problem is detected. A model experiencing drift may have affected hundreds or thousands of decisions before an alert triggers. Incident response planning must account for this retrospective impact assessment, tracing the effects back through prior decisions to understand the full scope of harm.
Model performance degradation occurs when accuracy drops below the declared threshold. Fairness drift arises when subgroup metrics breach declared thresholds, potentially affecting protected groups disproportionately. Data poisoning involves evidence of unauthorised training data modification that compromises model integrity. Adversarial exploitation covers evidence of adversarial inputs being used to manipulate system outputs. Privacy breach describes situations where model outputs reveal personal information from training data. Human oversight failure refers to evidence that system outputs were applied without the required human review, undermining the Human Oversight safeguards built into the deployment.
Each category should have clearly defined escalation paths and pre-approved containment actions so the response team can act without delay.
Incident response planning should define the channels through which each detection source reaches the response team and the criteria for classifying a detection event as a potential incident.
The triage phase uses a pre-defined matrix to classify severity and type. The matrix should distinguish between model-related incidents such as unexpected outputs, fairness violations, and accuracy degradation; data-related incidents such as data corruption, pipeline failure, and drift; infrastructure-related incidents such as outage, performance degradation, and security breach; and human oversight incidents such as operator error, automation bias, and interface failure. The triage outcome determines the response team composition, the escalation path, and the timeline for resolution.
Evidence is perishable: subsequent processing may overwrite the model's state, input data, feature values, and intermediate computations if they are not preserved immediately. The incident response plan must include immediate evidence preservation procedures that are triggered automatically at the point of detection.
The Technical SME tests the snapshot mechanism periodically as part of disaster recovery testing to confirm it captures all required evidence. This verification ensures the preservation procedure works reliably when an actual incident occurs, rather than discovering gaps during a real crisis.
Eradication involves identifying and removing the root cause. This may mean correcting corrupted data, fixing a pipeline bug, retraining a drifted model, or patching a security vulnerability. The distinction between containment and eradication is important: containment stops the bleeding while eradication addresses the underlying cause.
Deployer notification forms part of recovery: deployers who were informed of the issue during containment must be updated on the resolution and any actions they need to take, such as reverting to normal oversight levels. The staged approach ensures that a fix which inadvertently introduces new problems is caught during canary deployment rather than affecting all users simultaneously.
The incident response team must coordinate with the Legal and Regulatory Advisor and AI Governance Lead to prepare the initial report within the required timeframe. This coordination means the Serious Incident Reporting process runs in parallel with the technical response rather than as a sequential follow-up. The Communications Lead manages deployer and authority notifications while the Technical Lead focuses on diagnosis and remediation.
The key limitation of a manual approach is the loss of automated alerting and escalation. The on-call person must actively monitor dashboards rather than receiving push notifications, which increases response time. For systems where minutes matter, incident management platforms with free or low-cost tiers are strongly advisable to close this gap.
CTO of Standard Intelligence. Leads platform engineering and contributes to the PIG series technical content.
