← All news and guides
AI governance

AI Incident Response: What Counts as an AI Incident and How to Report It

What counts as an AI incident, which reporting clocks it starts (CERT-In, DPDP, EU AI Act), what ISO 42001 expects, and what to build before one hits.

Mono Team ·

A red emergency stop button on a plain grey industrial control panel, lit by a single overhead lamp in a quiet room.

Short answer: an AI incident is any event where an AI system's behaviour, or an attack on it, causes or could cause harm, whether or not data was stolen. That includes a wrong or harmful output reaching customers, one tenant's data surfacing in another's answer, an agent taking an action nobody approved, and a model provider's outage or breach. Handle it with the process you already use for security incidents, plus an AI layer: a kill switch, AI-specific logs and a classification step that asks which reporting clocks have started. ISO/IEC 42001 asks for an incident process but sets no reporting deadline. Your legal clocks come from CERT-In (6 hours, in force since 2022), DPDP Rule 7 (from 13 May 2027) and, for EU high-risk systems, the AI Act.

India has no AI-specific incident reporting rule as of November 2026. This is an engineering guide, not legal advice. Counsel should approve the wording of every external notice.

What counts as an AI incident?

NIST's Generative AI Profile (NIST AI 600-1, July 2024) borrows a definition that is useful as a starting point. An AI incident is an "event, circumstance, or series of events where the development, use, or malfunction of one or more AI systems directly or indirectly contributes to" harm. The harms it lists include injury, disruption of critical infrastructure, violations of human rights or of legal obligations, and harm to property, communities or the environment.

For a SaaS team that is too broad to run a pager on, so turn it into a working list. These are the events we suggest treating as incidents:

EventExampleNormal breach runbook enough?
Data exposure through AI outputExample: A summary shows another customer's text, or a model repeats a secret from a promptNormal breach runbook enough?: Mostly. Add AI scoping: which prompts, outputs and indexes held the data
Tenant isolation failure in retrievalExample: A missing tenant filter lets the retriever return another tenant's documentsNormal breach runbook enough?: Mostly, plus a retrieval fix and a test
Prompt injection that triggers an actionExample: A document tells an agent to email data out or change a recordNormal breach runbook enough?: Partly. You also need the tool-call log and a way to roll the action back
Harmful or materially wrong outputExample: A confident wrong answer in a regulated workflow; abusive text sent to a customerNormal breach runbook enough?: No. Nothing was breached, but the harm is real
Unapproved agent actionExample: An agent deletes records, spends money or sends messages outside its briefNormal breach runbook enough?: No. Needs action logs, permissions review and rollback
Bias or unfair outcome foundExample: A scoring feature behaves differently for one group of peopleNormal breach runbook enough?: No. Needs evaluation data and a decision on suspending the feature
Silent model changeExample: A provider updates a model and your accuracy or safety behaviour shiftsNormal breach runbook enough?: No. Needs version pinning, regression tests and provider contact
Provider incidentExample: Your model provider reports a breach or an outageNormal breach runbook enough?: Partly. Your duties depend on whether your customers' data was involved

The last column is the point. A runbook built only around stolen data will not trigger for many of these. Add a plain rule: if the AI did something harmful or wrong, it is an incident even if no attacker was involved.

Near misses count too. NIST AI 600-1 suggests procedures to record and track reported errors, near-misses and negative impacts (MG-4.3-002).

How is an AI incident different from a data breach?

A data breach, in the DPDP Act's sense, is about unauthorised processing or loss of personal data. An AI incident is wider, because it covers harm with no personal data involved, such as a bad automated decision. The two overlap whenever personal data is exposed. Then run the breach process as well, not instead. We cover the clocks, notices and evidence for that case in our CERT-In and DPDP breach runbook. This guide covers what to add for the AI side.

Which reporting clocks can an AI incident start?

A utility cabinet with rows of old black breaker switches, one lowered below the others, under a bare overhead bulb.

Triage should ask one question for each row below. Most incidents start none of these clocks. Some start more than one.

SourceWhat it coversDeadlineStatus as of 20 November 2026
CERT-In Directions, 28 April 2022What it covers: Cyber incidents listed in Annexure I, including data breach, data leak and (item xx) attacks or malicious activity affecting systems and applications related to AI and MLDeadline: 6 hours from noticing or being toldStatus as of 20 November 2026: In force
DPDP Rule 7What it covers: Personal data breaches: notify each affected Data Principal and the Board without delay, then a detailed Board reportDeadline: Without delay; 72 hours for the detailed reportStatus as of 20 November 2026: From 13 May 2027
EU AI Act, Art. 73What it covers: Serious incidents involving high-risk AI systems, reported by the provider to the market surveillance authorityDeadline: 15 days from awareness; 2 days for a widespread infringement or critical-infrastructure disruption; 10 days for a deathStatus as of 20 November 2026: High-risk dates moved by the AI Omnibus (see below)
EU AI Act, Art. 26(5)What it covers: A deployer that identifies a serious incident informs the provider first, then othersDeadline: ImmediatelyStatus as of 20 November 2026: Same timing note as above
ISO/IEC 42001What it covers: Management system: incident communication, corrective actionDeadline: None setStatus as of 20 November 2026: Voluntary standard
Your contractsWhat it covers: Customer DPAs and MSAs often set notice windowsDeadline: Whatever you signedStatus as of 20 November 2026: Check each contract

Note that the CERT-In duty is about cyber incidents on its list. An AI system that merely gives a wrong answer is not on it. A prompt injection attack that exfiltrates data probably is, and counsel should confirm edge cases.

Does the EU AI Act's incident reporting apply to an Indian SaaS company?

It can. The Act covers providers that place AI systems on the EU market whether or not they are established in the Union, and it covers third-country providers and deployers where the output of the system is used in the EU (Art. 2(1)). An Indian SaaS vendor selling an AI feature to EU customers is usually a provider. The EU customer is the deployer.

Art. 73 applies only to high-risk AI systems. A support chatbot or a document summariser is usually not high-risk. A feature used for employment decisions, education or biometrics falls into Annex III and may be. That is the first thing to settle, and our guide to the EU AI Act for Indian SaaS companies covers the classification question.

Dates have moved. The AI Omnibus (Regulation (EU) 2026/1744, OJ 24 July 2026) entered into force on 27 July 2026. It moved the Annex III high-risk date from 2 August 2026 to 2 December 2027, and the Annex I date to 2 August 2028. The Omnibus moved deadlines; obligations remain. Ask counsel whether the Omnibus changed Art. 73 itself and which date to work to for your system.

A serious incident, under Art. 3(49), is an incident or malfunction of an AI system that leads directly or indirectly to death or serious harm to health, disruption of critical infrastructure, an infringement of fundamental-rights obligations, or serious harm to property or the environment. For an Annex III system, build the reporting path now, with a named person and a record of when you became aware.

What does ISO 42001 expect for AI incidents?

ISO/IEC 42001:2023 does not give you a clock. It gives you controls and a loop. The Annex A controls most relevant to incidents are A.8.4 (communication of incidents), A.8.3 (external reporting, so interested parties can report adverse impacts), A.6.2.8 (AI system recording of event logs) and A.6.2.6 (AI system operation and monitoring).

Clause 10.2 covers nonconformity and corrective action. When something goes wrong, you must control and correct it, find the causes and whether similar problems exist or could occur, take the action needed and review whether it worked, change the management system if needed, and keep documented evidence.

In practice, an auditor or buyer wants to see a closed incident with a root cause, a fix, proof the fix worked and a changed control. That is a post-incident review, which is covered below. ISO 42001 sets no reporting deadline; your legal clocks come from CERT-In, DPDP and (for EU high-risk) the AI Act. See our ISO 42001 readiness page. Only an accredited certification body can certify you. Mono provides readiness, not certification.

What should you build before the first AI incident?

Most of the pain in an AI incident comes from missing controls, not from a missing document. Five engineering items make the biggest difference:

  • A kill switch at three levels. A global flag for the AI feature, a per-feature flag and a per-tenant toggle. Test that each one works without a deploy. This is the fastest containment you have, and it is the first question a security reviewer asks.
  • AI request logs with version stamps. Log request ID, tenant, user, feature, model and version, prompt template version, retrieved document IDs and tool calls. Avoid writing raw prompts and outputs to ordinary application logs, because that creates a second uncontrolled copy of customer data. Keep them in a restricted store with its own retention.
  • Pinned model versions and regression tests. Pin versions where the provider allows it, and run a fixed evaluation set on every change, so a silent provider change shows up as a failing test.
  • Tool permissions and rollback for agents. List the actions each agent can take, give it the least access it needs, and keep an undo path for anything that changes data.
  • A tenant-aware way to find who was affected. A saved query that takes a time window and a feature, and returns affected tenants and users. It is the slow step in nearly every incident. Mono builds these controls with your team.

These map to the ISO 42001 controls above and to the Manage function of the NIST AI RMF (prioritise, respond, recover), which our NIST AI RMF guide for SaaS turns into records. Buyers probe them in security reviews too.

What does an AI incident record need to contain?

NIST AI 600-1 suggests setting a minimum set of criteria for incident reporting (GV-4.3-002): a system ID, title, reporter, system or source, date reported, date of incident, description, impacts and stakeholders impacted. That is a sound base. Add the fields an AI team needs on top:

FieldWhy it matters
T0: when anyone became awareWhy it matters: Starts the legal clocks. Record the source: alert, customer, vendor, employee
Feature, model and version, prompt template versionWhy it matters: Tells you whether a rollback or pin would have prevented it
Tenants and users affected, with the query usedWhy it matters: The notice list and the evidence for it
Personal data involved (yes, no, unknown)Why it matters: Decides whether the DPDP and breach track starts
On CERT-In list (yes, no, check)Why it matters: Decides whether the 6-hour clock is running
EU high-risk system (yes, no)Why it matters: Decides whether Art. 73 reporting is in play
Containment action and timeWhy it matters: Shows how fast you stopped it
Root cause and corrective actionWhy it matters: ISO 42001 clause 10.2 evidence
Notices sent, with time and channelWhy it matters: Proof for customers and regulators

What should the first hour of an AI incident look like?

A long shelf of identical closed grey binders with plain unmarked spines, softly lit, the far end out of focus.

Here is a response order we suggest. The steps and targets are our recommendation, not legal requirements. The legal numbers are the ones in the table above.

  • Record T0 and open the ticket. Anyone can raise it. Do not wait for certainty.
  • Contain first. Use the kill switch, per-tenant toggle or a model rollback. Preserve before you clean up.
  • Preserve evidence. Snapshot the AI request logs, retrieval logs, tool-call logs, configuration and model versions in force at the time. Write them to storage that nobody can edit.
  • Classify. Answer the questions in the incident record: personal data, CERT-In list, EU high-risk, customer contracts. Each yes starts a clock.
  • Notify on the clocks that started. CERT-In within 6 hours if it is a listed incident. Customers within contract windows. Users and the Board under Rule 7 from 13 May 2027, if personal data was involved.
  • Scope. Use the saved query to list affected tenants and users.
  • Fix the cause, not only the symptom. A missing tenant filter, a permissive tool, an unpinned model.
  • Review. Hold a short post-incident review within a few days, with owners and dates for each corrective action.

Decide in advance who leads, who can switch the feature off, who talks to customers and who approves external wording. Include an engineer who reads AI logs.

What if the incident starts at your AI model provider?

Provider incidents are a different problem, because you learn about them late and have limited control. NIST AI 600-1 (GV-6.2) suggests incident response plans for third-party generative AI, with defined ownership, regular rehearsal and vendor contract terms that cover incident response and response times.

Public contract terms show why that matters. OpenAI's data processing addendum (updated 1 December 2025) commits to notifying a personal data breach "without undue delay", with no hour figure, and gives 30 days to object to subprocessor changes. Tighter hour limits are negotiated, and the Common Paper AI Addendum template has no incident-notice clause, so that comes from the DPA.

The gap matters for you. A vendor can legitimately tell you late, and your own customers may expect notice within a day or two. Know your vendors' notice terms, keep a list of every subprocessor in the AI data path, and have your lawyers review the clauses. Our AI security questionnaire guide shows how buyers ask about this.

How do you test an AI incident process?

Run a tabletop each quarter with AI-shaped scenarios. Four that work well:

  • Cross-tenant leak. A customer reports seeing another company's text in a summary. Can you contain it in minutes and name the affected tenants within hours?
  • Prompt injection. A document makes an agent send an export outside the company. What do the tool-call logs show, and what can you roll back?
  • Provider outage or breach. Your model vendor posts a notice at 7 pm on a Friday. Who reads it, and who tells customers?
  • Harmful output. A feature produces abusive or clearly wrong text for a regulated workflow. Who decides to switch it off, and who writes the customer note?

Time every step, record the gaps as fixes with an owner and a date, and keep the notes. They are the evidence an ISO 42001 auditor or an enterprise buyer will ask for.

How do you answer "do you have an AI incident response process?" in a questionnaire?

Say what exists, in specific terms, and attach evidence. A strong answer names the AI section of your incident plan, the kill switch, the log fields, the owner and the date of the last tabletop. A weak answer says the plan "covers all incidents". If you lack a kill switch or AI logs, say so and give a date.

If your next deal is waiting on this, we help teams answer AI security questionnaires and build the controls behind the answers. We do not certify anyone, and this is not legal advice.

Sources: NIST, Generative Artificial Intelligence Profile, NIST AI 600-1, 26 July 2024 (A.1.8, GV-4.3, GV-6.2, MG-4.3). NIST AI RMF 1.0 core functions (airc.nist.gov). CERT-In Directions under Section 70B(6) of the IT Act, 2000, No. 20(3)/2022-CERT-In, 28 April 2022, direction (ii) and Annexure I. Digital Personal Data Protection Rules, 2025, Rule 7. Regulation (EU) 2024/1689 (AI Act), Articles 2, 3(3), 3(4), 3(49), 26(5) and 73, and the European Commission's AI Act page on dates and the AI Omnibus (Regulation (EU) 2026/1744, OJ 24 July 2026). ISO/IEC 42001:2023, clause 10.2 and Annex A controls A.6.2.6, A.6.2.8, A.8.3 and A.8.4. OpenAI Data Processing Addendum and Services Agreement (December 2025); Anthropic Commercial Terms (17 June 2025); Common Paper AI Addendum v1.0. Last reviewed 20 November 2026.

FAQ

What counts as an AI incident?

An AI incident is an event where the development, use or malfunction of an AI system directly or indirectly causes or contributes to harm. In a SaaS product that includes data shown to the wrong customer, a prompt injection attack, an agent acting without approval, harmful or materially wrong output, a harmful bias finding, a silent model change and a model provider incident. It does not need a data breach.

How long do you have to report an AI incident in India?

India has no AI-specific reporting rule as of November 2026. If the incident is on CERT-In's list of cyber incidents, which includes attacks affecting AI and machine learning systems, you must report within 6 hours of noticing it. From 13 May 2027, a personal data breach also triggers DPDP Rule 7: notify affected people and the Board without delay, then file a detailed report within 72 hours.

Does the EU AI Act require serious incident reporting?

Yes, for high-risk AI systems. Under Article 73, the provider reports a serious incident to the market surveillance authority within 15 days of becoming aware, within 2 days for a widespread infringement or critical-infrastructure disruption, and within 10 days for a death. The AI Omnibus moved the Annex III high-risk date to 2 December 2027. Ask counsel which date applies to your system.

Does ISO 42001 set a deadline for reporting AI incidents?

No. ISO/IEC 42001 asks you to communicate incidents, record events and run corrective action under clause 10.2, but it sets no time limit. Your deadlines come from CERT-In, DPDP and, for EU high-risk systems, the AI Act, plus whatever your customer contracts say.

How is an AI incident response plan different from a breach runbook?

A breach runbook assumes data was exposed. An AI incident plan also covers harm with no breach, such as wrong or harmful output, an unapproved agent action or a provider's model change. It adds a kill switch, AI request logs with model and prompt versions, a near-miss log and a classification step that checks which reporting clocks have started. Run both when personal data is involved.

What should an AI incident record include?

At least a system ID, title, reporter, date of the incident and date reported, description, impacts and stakeholders affected, which is the minimum NIST AI 600-1 suggests. Add the time anyone became aware, model and prompt versions, affected tenants and the query that found them, whether personal data is involved, the containment action and time, root cause, corrective action and notices sent.

Get an AI incident process buyers can see in writing.