MANAGE
This content is not available in your language yet.
Jak prioritizovat rizika, implementovat mitigace a řídit incidenty
NIST AI RMF funkce: MANAGE Subkategorie: MG-1 až MG-4
1. Přehled funkce MANAGE
MANAGE funkce převádí výsledky z MAP a MEASURE do konkrétních akcí.
Cíle MANAGE
| Cíl | Popis |
|---|---|
| Prioritizace | Seřadit rizika podle závažnosti |
| Mitigace | Implementovat kontrolní opatření |
| Incidenty | Řídit AI incidenty |
| Improvement | Kontinuálně zlepšovat |
2. MG-1: Risk Prioritization
2.1 Risk scoring
Risk Score = Likelihood × Impact × Detectability
| Faktor | 1 (Low) | 2 (Medium) | 3 (High) | 4 (Critical) |
|---|---|---|---|---|
| Likelihood | Rare | Possible | Likely | Almost certain |
| Impact | Minor | Moderate | Major | Severe |
| Detectability | Easy | Moderate | Difficult | Very difficult |
Score interpretace:
| Score | Level | Action |
|---|---|---|
| 1-8 | Low | Accept or monitor |
| 9-16 | Medium | Mitigate within quarter |
| 17-32 | High | Mitigate within month |
| 33-64 | Critical | Immediate action |
2.2 Risk register
## AI RISK REGISTER
| ID | Risk | System | L | I | D | Score | Status | Due ||----|------|--------|---|---|---|-------|--------|-----|| R001 | Bias in hiring | AI-001 | 3 | 4 | 2 | 24 | Mitigating | 2025-02 || R002 | Data privacy | AI-002 | 2 | 3 | 3 | 18 | Accepted | - || R003 | Hallucination | AI-003 | 4 | 3 | 3 | 36 | Critical | 2025-01 |2.3 Risk treatment options
| Option | Kdy použít | Příklad |
|---|---|---|
| Avoid | Nepřijatelné riziko | Nepoužívat AI pro daný účel |
| Mitigate | Snížitelné riziko | Implementovat kontroly |
| Transfer | Přenositelné riziko | Pojištění, SLAs |
| Accept | Akceptovatelné zbytkové riziko | Dokumentovat a monitorovat |
3. MG-2: Risk Mitigation
3.1 Control framework
3.2 Kontroly dle rizika
Pro Bias/Fairness rizika
| Kontrola | Typ | Implementace |
|---|---|---|
| Diverse training data | Preventive | Data curation process |
| Bias testing pre-deployment | Preventive | Mandatory fairness check |
| Fairness monitoring | Detective | Continuous metric tracking |
| Human review for decisions | Detective | Sample review process |
| Appeal mechanism | Corrective | User-facing appeal process |
| Retraining pipeline | Corrective | Triggered by drift |
Pro Privacy rizika
| Kontrola | Typ | Implementace |
|---|---|---|
| Data minimization | Preventive | Only necessary data |
| Anonymization/pseudonymization | Preventive | Pre-processing pipeline |
| Access controls | Preventive | RBAC, least privilege |
| PII detection in outputs | Detective | Content scanning |
| Data breach response | Corrective | Incident response plan |
Pro Safety rizika (GAI)
| Kontrola | Typ | Implementace |
|---|---|---|
| Content filters | Preventive | Input/output filtering |
| System prompts | Preventive | Safety instructions |
| Rate limiting | Preventive | Abuse prevention |
| Output monitoring | Detective | Toxicity scoring |
| Kill switch | Corrective | Immediate disable capability |
| User blocking | Corrective | Ban abusive users |
3.3 Human oversight
Levels of human oversight:
| Level | Popis | Kdy použít |
|---|---|---|
| Human-in-the-loop | Člověk schvaluje každé rozhodnutí | High-risk, critical decisions |
| Human-on-the-loop | Člověk monitoruje a může zasáhnout | Medium-risk, batch decisions |
| Human-in-command | Člověk nastavuje pravidla, AI operuje | Low-risk, high-volume |
Implementace:
## HUMAN OVERSIGHT DESIGN
### 1. Decision Point- Co AI doporučuje vs. rozhoduje?- Kdy je vyžadován human review?
### 2. Escalation Criteria| Trigger | Akce ||---------|------|| Confidence < 80% | Human review || High-impact decision | Mandatory approval || Edge case detected | Escalate to expert || User requests | Always allow |
### 3. Override Capability- Kdo může overridnout AI?- Jak je override logován?- Jaký je feedback loop do modelu?
### 4. Appeal Process- Jak může dotčená osoba podat odvolání?- Kdo rozhoduje o odvolání?- Jaké jsou lhůty?4. MG-3: Incident Management
4.1 AI Incident definice
Co je AI incident:
“Událost nebo série událostí, kde vývoj, použití nebo porucha AI systému přímo či nepřímo přispívá ke škodě na zdraví, narušení kritické infrastruktury, porušení lidských práv nebo škodě na majetku/komunitách/životním prostředí.”
4.2 Incident kategorie
| Kategorie | Příklady | Severity |
|---|---|---|
| Safety | Fyzická újma, nebezpečný obsah | Critical |
| Privacy | Data breach, PII leak | Critical |
| Bias/Discrimination | Unfair outcomes, discrimination | High |
| Misinformation | Hallucinations with impact | High |
| Security | Adversarial attack, system compromise | Critical |
| Performance | Major accuracy degradation | Medium |
| Availability | System outage | Medium-High |
4.3 Incident response process
4.4 Incident response SLAs
| Severity | Detection → Triage | Triage → Containment | Resolution | Communication |
|---|---|---|---|---|
| P1 Critical | 15 min | 1 hour | ASAP | Immediate |
| P2 High | 1 hour | 4 hours | 24 hours | 4 hours |
| P3 Medium | 4 hours | 24 hours | 1 week | 24 hours |
| P4 Low | 24 hours | 1 week | 1 month | As needed |
4.5 Regulatory notification
| Regulace | Trigger | Lhůta | Komu |
|---|---|---|---|
| GDPR | Personal data breach | 72 hodin | ÚOOÚ |
| NIS2 | Significant incident | 24h early warning, 72h notification | NÚKIB |
| DORA | Major ICT incident | 4 hodiny | ČNB |
| EU AI Act | Serious incident | 15 dní (nebo ihned) | Market surveillance |
4.6 Incident report template
# AI INCIDENT REPORT
## Metadata| Field | Value ||-------|-------|| Incident ID | INC-2025-XXXX || System | || Detected | YYYY-MM-DD HH:MM || Resolved | YYYY-MM-DD HH:MM || Severity | P1/P2/P3/P4 || Category | Safety/Privacy/Bias/Security/Performance |
## Summary[1-2 sentence summary]
## Timeline| Time | Event ||------|-------|| | Detection || | Triage || | Containment || | Resolution |
## Impact- Users affected:- Data affected:- Business impact:- Reputation impact:
## Root Cause[Description of underlying cause]
## Resolution[What was done to fix]
## Preventive Actions| Action | Due ||--------|-----|| | |
## Lessons Learned[Key takeaways]
## Regulatory Notifications| Regulator | Notified | Reference ||-----------|----------|-----------|| | | |5. MG-4: Continuous Improvement
5.1 Improvement cycle
5.2 Review cadence
| Review | Frekvence | Scope | Účastníci |
|---|---|---|---|
| Operational | Weekly | Active incidents, metrics | AI Ops team |
| Tactical | Monthly | Risk register, controls | AI Risk Officer + Owners |
| Strategic | Quarterly | Governance, trends | Leadership |
| Annual | Yearly | Full program review | Board |
5.3 Improvement sources
| Zdroj | Typ inputů | Jak získat |
|---|---|---|
| Incidents | Lessons learned | Post-mortems |
| Metrics | Performance trends | Dashboards |
| Audits | Findings, recommendations | Internal/external audits |
| Feedback | User complaints, suggestions | Feedback channels |
| Regulatory | New requirements | Regulatory monitoring |
| Industry | Best practices | Conferences, publications |
5.4 Decommissioning
Kdy decommissionovat AI systém:
- Performance trvale pod thresholdem
- Business need již neexistuje
- Bezpečnostní riziko nelze mitigovat
- Regulatorní požadavky nelze splnit
- Lepší alternativa je dostupná
Decommissioning checklist:
- Notifikovat stakeholdery
- Zajistit data retention/deletion dle požadavků
- Aktualizovat inventuru
- Dokumentovat lessons learned
- Archivovat relevantní dokumentaci
- Potvrdit compliance s data retention policies
6. Implementační checklist
Fáze 1: Risk Treatment (Týden 1-2)
- Vytvořit risk register
- Definovat scoring metodologii
- Přiřadit vlastníky rizik
- Navrhnout mitigační plány
Fáze 2: Controls (Týden 3-4)
- Implementovat preventivní kontroly
- Nastavit detektivní kontroly
- Připravit korektivní mechanismy
- Dokumentovat human oversight
Fáze 3: Incident Response (Týden 5-6)
- Definovat incident process
- Přiřadit role a odpovědnosti
- Nastavit komunikační kanály
- Provést tabletop exercise
Fáze 4: Continuous (Ongoing)
- Pravidelné reviews
- Incident post-mortems
- Metric-driven improvements
- Regulatory updates
7. Šablony
| Šablona | Účel | Lokace |
|---|---|---|
| Risk Register | Evidence a tracking rizik | /templates/open/ |
| Incident Report | Dokumentace incidentů | /templates/open/ |
| Post-mortem Template | Lessons learned | /templates/open/ |
| Decommissioning Checklist | Ukončení AI systému | /templates/open/ |
Pokračujte na GAI Rizika pro specifika generativní AI.
AI-Native Entry Framework | CC BY-NC-SA 4.0