Quick answer
CERT-In's directions of 28 April 2022, issued under section 70B(6) of the IT Act, apply to AI systems like any other ICT system: report specified cyber incidents within 6 hours of noticing them, keep ICT system logs securely for a rolling 180 days within Indian jurisdiction, and synchronise clocks to NIC or NPL NTP servers. DPDP adds one-year log retention from 13 May 2027.
- CERT-In's 28 April 2022 directions apply to AI infrastructure; there is no AI exemption for LLM gateways, model APIs or agent tool calls.
- Specified cyber incidents must be reported to CERT-In within 6 hours of noticing them.
- ICT system logs must be kept securely for a rolling 180 days within Indian jurisdiction.
- DPDP Rules 6(1)(e) and 8(3) require at least one year of logs from 13 May 2027, so the longer period governs design.
- Prompt injection or model compromise that exposes personal data can trigger CERT-In and DPDP reporting at the same time.
Who this is for: Platform and SRE teams running AI workloads, CISOs and security operations centres, AI SaaS and API providers operating in India, Cloud, GPU and data centre providers, Incident response and compliance leads
What do CERT-In's directions require?
The Indian Computer Emergency Response Team (CERT-In) issued directions on 28 April 2022 under section 70B(6) of the IT Act, 2000. They apply to service providers, intermediaries, data centres, body corporates and government organisations. The core requirements:
| Requirement | What it says |
|---|---|
| Incident reporting | Report specified cyber incidents to CERT-In within 6 hours of noticing them or being notified |
| Log retention | Maintain logs of ICT systems securely for a rolling 180 days, within Indian jurisdiction, and provide them to CERT-In on request |
| Clock synchronisation | Synchronise ICT system clocks to the NIC or NPL NTP servers, or to sources traceable to them |
| Point of contact | Designate a point of contact to interface with CERT-In |
| Subscriber records | VPN service providers, cloud service providers and data centres keep specified subscriber and customer records for 5 years |
The directions list the categories of incidents that must be reported — for example, compromise of systems, unauthorised access, data breaches and attacks on servers and applications. Check the full list on cert-in.org.in, along with CERT-In's FAQs, which clarify how the directions are applied.
None of this is AI-specific, and that is the point. AI systems are ICT systems, so the directions apply to them without any further rule.
Applying the directions to AI systems
Modern AI stacks add components that security teams have not always logged. Each one is an ICT system.
Components in scope
- LLM gateway or proxy — the service routing prompts to internal or external models.
- Model inference endpoints — self-hosted models on GPUs, or managed endpoints.
- Vector databases and retrieval services — where embeddings and source chunks live.
- Agent orchestrators and tool calls — the layer that lets a model call APIs, run code or send messages.
- Fine-tuning and training pipelines — data loaders, training jobs, model registries.
- Evaluation and observability tools — trace stores that capture prompts and outputs.
AI-flavoured incidents that may be reportable
- unauthorised access to a model endpoint or gateway credentials;
- data exfiltration through a model, for instance via prompt injection that makes a RAG system return restricted documents;
- compromise of an agent's tool credentials leading to unauthorised actions;
- tampering with training data or model weights;
- malicious code introduced through a model or dependency supply chain.
Whether a specific event falls within a listed category is a judgement call. When in doubt, treat a security compromise of an AI component the same way you would treat a compromise of a web application, and assess it against the CERT-In list within the 6-hour window.
Designing AI logs for 180 days and one year
CERT-In asks for 180 days of logs in India. DPDP Rules ask for at least one year from 13 May 2027: Rule 6(1)(e) for detection and investigation, and Rule 8(3) for personal data, traffic data and processing logs for Seventh Schedule purposes. Design for the longer period.
What to log at each layer
| Layer | Minimum fields |
|---|---|
| LLM gateway | timestamp (NTP-synced), caller identity, application, model and version, token counts, policy decisions, request ID |
| Inference endpoint | request ID, latency, errors, auth result, source IP |
| Retrieval | query ID, documents retrieved (IDs, not content), access-control decision |
| Agent tool calls | tool name, arguments hash or redacted arguments, result status, acting identity |
| Admin actions | model deployment, prompt template changes, key rotation |
Prompts and outputs
Full prompts and responses often contain personal data. Options:
- store them encrypted with restricted access, as Rule 6(1)(a) contemplates;
- store redacted or tokenised versions, keeping the mapping in a separate vault;
- keep full content for a shorter period only if another law does not require it, while keeping metadata logs for the full retention period.
Location
The 180-day logs must be within Indian jurisdiction. If your primary AI region is abroad, ship a log copy to an Indian region. See LLM logging under DPDP.
Building a 6-hour reporting workflow for AI incidents
Six hours is short. The workflow has to be pre-built.
- Detection. Alerts from the LLM gateway (unusual volume, policy-violation spikes, new callers), endpoint authentication failures, and data-loss rules on model outputs.
- Triage (first hour). On-call engineer confirms the event, records the time of noticing and opens an incident ticket tagged "AI".
- Classification (hours 1–3). Security lead checks the event against CERT-In's reportable categories and decides whether personal data is involved.
- Reporting (by hour 6). The designated point of contact submits the report through CERT-In's prescribed channel with available details. Updates follow as the investigation proceeds.
- Parallel DPDP track. If personal data is affected, from 13 May 2027 Rule 7 requires intimation without delay to each affected Data Principal and to the Data Protection Board, with a detailed report within 72 hours.
Pre-build these
- a report template with fields pre-filled (entity, contact, systems);
- a decision tree for "reportable to CERT-In?" and "personal data breach?";
- a 24x7 rota that includes someone who understands the AI stack.
For the DPDP side, see the breach response playbook.
Cloud, GPU and model API providers
CERT-In's directions place specific record-keeping duties on VPN service providers, cloud service providers and data centres: subscriber and customer records must be kept for 5 years. GPU cloud and model hosting providers serving Indian customers should assess whether they fall within these categories.
For AI builders using third-party providers:
- Contract for logs. Ensure the provider can supply logs relevant to your tenancy within the reporting window and supports retention in Indian jurisdiction.
- Contract for incident notification. Require notification fast enough that you can meet your own 6-hour clock.
- Clock sync. Confirm the provider's timestamps are reliable and convertible to IST and NTP-traceable time.
- Foreign model APIs. Calls to a model API hosted abroad still generate logs on your side; keep your gateway logs in India.
DPDP overlap
Under DPDP, you remain responsible for processing done by your Data Processors (section 8(1)). Rule 6(1)(f) requires security clauses in processor contracts, and Rule 8(3) Illustration Case 2 makes the fiduciary ensure a cloud provider retains data and logs for at least one year. See AI vendor due diligence.
Common mistakes and next steps
- Logging only the application, not the model layer. Gateways and tool calls are where AI incidents show up.
- Logs only in a foreign region. CERT-In requires Indian jurisdiction for the 180-day logs.
- Unsynchronised clocks across GPU nodes. Timeline reconstruction fails without NTP-traceable time.
- Assuming prompt injection is not a security incident. If it leads to unauthorised access or exfiltration, it probably is.
- Running CERT-In and DPDP responses separately. One incident, two reports, one evidence base.
- Deleting logs at 180 days. From 13 May 2027, DPDP requires at least one year.
Before 13 May 2027
Extend retention to at least one year, add AI components to the asset register and SIEM, rehearse a joint CERT-In and DPDP incident tabletop, and update vendor contracts. These steps also support RBI's FREE-AI ask for AI incident reporting in financial services.
Further reading on suhasbhairav.com
Before 13 May 2027
Compliance checklist
Register AI components as ICT assets
Add gateways, inference endpoints, vector stores and agent orchestrators to the asset register and SIEM.
CERT-In Directions, 28 Apr 2022
Synchronise clocks
Point all AI nodes to NIC or NPL NTP servers or traceable sources.
CERT-In Directions
Keep logs in India
Retain AI system logs within Indian jurisdiction for at least 180 days.
CERT-In Directions
Extend retention to one year
Retain logs and associated personal data for at least one year from processing, then erase.
DPDP Rule 6(1)(e), Rule 8(3)
Protect prompt and output logs
Encrypt, mask or tokenise personal data in prompts and outputs and restrict access.
DPDP Rule 6(1)(a)-(b)
Designate a CERT-In point of contact
Name and publish internally the person who files reports.
CERT-In Directions
Pre-build the 6-hour workflow
Prepare templates, decision trees and a 24x7 rota that covers AI systems.
CERT-In Directions
Align with DPDP breach reporting
Map reportable events to Rule 7 intimations and the 72-hour Board report.
DPDP Rule 7
Update cloud and model vendor contracts
Require log access, incident notification and one-year retention from providers.
DPDP Rule 6(1)(f)
Key dates
CERT-In issues directions under section 70B(6) of the IT Act
DPDP Rules notified
DPDP security, breach and one-year log retention obligations apply
Answers
Frequently asked questions
Do CERT-In rules apply to AI systems?
Do CERT-In rules apply to AI systems?
Yes. CERT-In's directions of 28 April 2022 apply to ICT systems operated by service providers, intermediaries, data centres, body corporates and government organisations. LLM gateways, model endpoints, vector databases and agent tool layers are ICT systems, so incident reporting, log retention and clock synchronisation duties apply.
How quickly must AI security incidents be reported to CERT-In?
How quickly must AI security incidents be reported to CERT-In?
Within 6 hours of noticing the incident or being notified of it, if it falls within the categories of cyber incidents listed in the directions. Report with the information available and update as investigation continues. If personal data is involved, DPDP Rule 7 reporting runs in parallel from 13 May 2027.
How long must AI logs be kept in India?
How long must AI logs be kept in India?
CERT-In requires ICT system logs for a rolling 180 days within Indian jurisdiction. From 13 May 2027, DPDP Rules 6(1)(e) and 8(3) require logs and associated personal data to be kept for at least one year. Design retention for one year to satisfy both.
Can AI logs be stored outside India?
Can AI logs be stored outside India?
You can keep copies abroad, but CERT-In requires the 180-day logs to be maintained within Indian jurisdiction. Many teams ship a copy of gateway and security logs to an Indian region while running models elsewhere. Check sectoral rules too, such as RBI's payment data localisation.
Is prompt injection a reportable cyber incident?
Is prompt injection a reportable cyber incident?
It can be. If prompt injection leads to unauthorised access, data exfiltration or unauthorised actions through an agent's tools, it may fall within CERT-In's reportable categories. Assess it against the official list. If personal data is exposed, it is also a personal data breach under DPDP.
Do I need to log full prompts for CERT-In?
Do I need to log full prompts for CERT-In?
The directions require ICT system logs, not necessarily full prompt content. Metadata logs such as caller identity, model, timestamps and policy decisions are usually essential. Full prompts often contain personal data, so if you keep them, protect them under DPDP Rule 6 with encryption or masking and access control.
Legal references and sources
- CERT-In Directions under section 70B(6) of the IT Act — Directions of 28 April 2022 (official text)
- Digital Personal Data Protection Rules, 2025 — Rule 6(1)(c) and 6(1)(e) (official text)
- Digital Personal Data Protection Rules, 2025 — Rule 7 and Rule 8(3) (official text)
This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.
About the author

Suhas Bhairav
Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.
Related guides
- AI & personal dataLLM Logging and Observability Under DPDP's One-Year RuleDPDP asks you to minimise data yet keep processing logs for a year. A tiered design for LLM traces resolves the tension: log less, mask early, lock access, then erase on schedule.
- DPDP Rules 2025DPDP One-Year Log Retention: What to Keep and for How LongTwo DPDP rules require one year of retention: Rule 6(1)(e) for security logs and Rule 8(3) for personal data, traffic data and processing logs. What to keep, where, and how to protect it.
- Templates & playbooksDPDP Breach Response Playbook: The First 72 HoursAn hour-by-hour DPDP breach response playbook for the first 72 hours: triage, CERT-In's 6-hour report, Rule 7 intimations to users and the Board, and evidence.
- DPDP Rules 2025DPDP Data Breach Notification: The 72-Hour Rule ExplainedRule 7 requires every personal data breach to be reported to affected individuals and the Data Protection Board, with a detailed Board report within 72 hours. No materiality threshold applies.
- AI & personal dataAI Agents and Personal Data: DPDP Controls for Agentic SystemsAI agents read, write and share personal data through tools. These are the DPDP controls for tool scopes, approvals, memory, action logs and breach response.
- India AI & sector rulesData Localisation in India: Where Can You Host AI Workloads?Where you can legally host AI workloads that touch Indian personal data: DPDP's negative list, Rule 13(4) for SDFs, RBI payment data, CERT-In logs and sectoral rules, in one decision table.