What it means
- Recognise, record, respond to and learn from AI incidents, not just security breaches. AI incidents include:
- harmful or unlawful outputs, deepfakes made with your tools,
- biased or wrong decisions at scale, silent model degradation,
- prompt injection or data leakage through AI,
- agents taking unintended actions.
- Report to regulators, law enforcement and affected people when rules require it, and put right harm to customers.
- ISO 27001 reuse: Strong. The ISO 27001 incident management process, tooling and evidence handling carry over directly. New parts: AI incident categories, non-security harms (bias, wrong decisions), and AI-specific regulator channels.
Why it matters
- RBI FREE-AI asks regulated entities to record and report AI incidents promptly through the channel regulators set, using the indicative format in Annexure VI, and to compensate harmed customers.
- The 2026 IT Rules require reporting offences the law requires to be reported (for example under BNSS or POCSO).
- MeitY advisories tell platforms to inform users that legal violations will be reported to law enforcement.
- CERT-In's 6-hour cyber incident reporting still applies when an AI incident is also a cyber incident; DPDP breach notification applies when personal data is involved.
- The India AI Governance Guidelines plan a national AI incidents database.
Checklist
- Extend the incident policy with an AI incident definition, AI-specific categories and severity criteria.
- Add an "AI involved?" field to the incident ticket form.
- Map reporting duties and clocks: CERT-In (6 hours), DPDP Board, RBI or sector regulator channel, law enforcement for reportable offences, affected customers.
- Create AI playbooks: harmful output, data leakage via LLM, prompt injection, model drift causing wrong decisions, deepfake misuse of your tool.
- Include the kill switch and fallback steps in playbooks (SC-AI-026 Human oversight and override, SC-AI-033 Business continuity and fallback for AI).
- Preserve evidence: prompts, outputs, model version, logs (SC-AI-032 Logging and record keeping).
- Define how harmed customers are identified, contacted and compensated.
- Hold a post-incident review; feed fixes into risk assessment and change management (SC-AI-005 AI risk assessment, SC-AI-019 Change management and approval gates).
- Run at least one AI incident tabletop exercise a year.
Evidence examples
- Incident management policy with the AI addendum.
- AI incident register with severity, timeline, root cause and closure.
- Copies of regulator or CERT-In submissions with timestamps.
- AI playbooks and the latest tabletop exercise report.
- Post-incident review notes and resulting corrective actions.
- Customer compensation records for AI-caused harm.
Common pitfalls
- Only security breaches count as incidents; a biased model affecting thousands of loans never gets logged.
- No one knows which regulator to tell or within what time.
- Logs needed for investigation expire before anyone looks.
- Staff unsure whether a strange chatbot answer is "an incident", so nobody reports it.
Satisfies
- VoluntaryISO/IEC 42001
- BindingEU AI Act
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- SupervisoryNYDFS AI Cybersecurity Letter 2024
- BindingTexas TRAIGA
- BindingKorea AI Basic Act
- SupervisoryOSFI E-23
- SupervisoryMeitY AI and Deepfake Advisories
- VoluntaryRBI FREE-AI
Also relevant to
- India AI Governance Guidelinesguidance mapping (national AI incidents database; no obligation IDs)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.3.3 reporting of concerns
- A.8.3 external reporting
- A.8.4 communication of incidents
- 10.2 nonconformity and corrective action
ISO/IEC 27001
- A.5.24 incident management planning and preparation
- A.5.25 assessment and decision on events
- A.5.26 response to incidents
- A.5.27 learning from incidents
- A.5.28 collection of evidence
- A.6.8 event reporting
SOC 2
- CC2.3
- CC7.3
- CC7.4
- CC7.5
Key takeaways
- Widen "incident" to cover AI harms that are not security breaches.
- Know every reporting clock: CERT-In, DPDP, RBI or sector regulator, law enforcement.
- Reuse your ISO 27001 incident process; add AI playbooks and categories.