Glossary

DPDP and AI governance glossary.

Plain-English definitions of the terms in the DPDP Act and Rules, and of the standards and risks that enterprise buyers ask AI companies about. Each entry says what it means for the people who build the systems.

Quoted from the ActWhat it means for engineers
Glossary

Five groups of terms

DPDP Act terms10 terms
DPDP Rules and duties6 terms
Words the DPDP Act does not define3 terms
AI governance and assurance terms11 terms
AI security and engineering terms11 terms
Checked 20 November 2026Sources named on each entry
Not legal advice

DPDP Act terms.

These come from section 2 of the Digital Personal Data Protection Act, 2023. The quoted wording is the Act’s own.

Personal data

“Any data about an individual who is identifiable by or in relation to such data.”

In practice: If a record can be tied to a person, alone or with other fields, treat it as personal data. A recorded voice or face of an identifiable person counts. Put the logs, exports and copies made from your databases in your data map too.

Source: DPDP Act, s.2(t). Start with a data map

Data Principal

“The individual to whom the personal data relates.” For a child, this includes the parents or lawful guardian. For a person with a disability, it includes her lawful guardian acting on her behalf.

In practice: This is your user, customer or employee. Your rights flows (access, correction, erasure, grievance) are built around this person. Where the person is a child, the flow has to reach the parent or guardian too.

Source: DPDP Act, s.2(j). Rights request workflow

Data Fiduciary

“Any person who alone or in conjunction with other persons determines the purpose and means of processing of personal data.”

In practice: If you decide why and how personal data is used, you are the fiduciary, and the duties in the Act sit with you. You stay responsible when a vendor does the work for you (s.8(1)). Engineers should know which of their systems make that decision, and which only follow it.

Source: DPDP Act, s.2(i). The DPDP Act explained

Data Processor

“Any person who processes personal data on behalf of a Data Fiduciary.”

In practice: The fiduciary needs a valid contract with each processor (s.8(2)), and the security duty covers processing a processor does on its behalf (s.8(5)). A software company can be a processor for its customers’ data and a fiduciary for its own. The facts of each flow decide which.

Source: DPDP Act, s.2(k), s.8. SaaS as a data processor

Processing

“A wholly or partly automated operation or set of operations performed on digital personal data.” It includes collection, recording, storage, use, sharing, disclosure by transmission, restriction, erasure and destruction.

In practice: Almost everything your systems do with the data is processing: logging it, backing it up, sending it to a model API, and deleting it. Each of those needs a purpose and a safeguard. The definition covers digital personal data.

Source: DPDP Act, s.2(x). DPDP and AI tools

Specified purpose

“The purpose mentioned in the notice given by the Data Fiduciary to the Data Principal.”

In practice: The notice fixes the purpose, and the purpose limits what you may do with the data. Record the purpose against each data store or field, so you can show what it is for and when it stops being needed. Under s.8(7), data is erased once the specified purpose is no longer served, unless a law requires you to keep it.

Source: DPDP Act, s.2(za), s.8(7). Erasure automation

Child

“An individual who has not completed the age of eighteen years.”

In practice: Before processing a child’s personal data, a fiduciary must get verifiable consent from the parent or lawful guardian (s.9(1)). The Act also bars tracking, behavioural monitoring and targeted advertising directed at children (s.9(3)), subject to the exemptions in s.9(4) and s.9(5). That is a product decision as well as a legal one: an age check, a parent step, and ad and analytics settings that differ by age.

Source: DPDP Act, s.2(f), s.9. Verifiable parental consent

Personal data breach

“Any unauthorised processing of personal data or accidental disclosure, acquisition, sharing, use, alteration, destruction or loss of access to personal data, that compromises the confidentiality, integrity or availability of personal data.”

In practice: A breach is wider than a hack. A mis-sent export, a public storage bucket or a ransomware lockout can all qualify. Under Rule 7, you tell affected people and the Board without delay, and send the Board a detailed report within 72 hours. CERT-In has its own, shorter, 6 hour clock for listed incidents. Build one runbook for both.

Source: DPDP Act, s.2(u); DPDP Rules, r.7. Breach: 72 hours and CERT-In

Significant Data Fiduciary (SDF)

A fiduciary or class of fiduciaries “notified by the Central Government under section 10.” The government decides on the basis of factors that include the volume and sensitivity of personal data processed, risk to Data Principals’ rights, and public order.

In practice: As of November 2026, no Significant Data Fiduciary has been notified. If you are notified, the extra duties in s.10(2) apply: a Data Protection Officer based in India, an independent data auditor, periodic Data Protection Impact Assessments and periodic audits. Mandatory audits apply only to these fiduciaries. Nothing in the Act or Rules creates a general DPDP certificate.

Source: DPDP Act, s.2(z), s.10; DPDP Rules, r.13. Significant Data Fiduciary readiness

DPDP Rules and duties.

Terms from the DPDP Rules, 2025 (G.S.R. 846(E)) and the duties most engineering teams meet first. Core duties start on 13 May 2027.

Reasonable security safeguards (Rule 6)

Section 8(5) requires a Data Fiduciary to protect personal data in its possession or under its control, including processing by its Data Processors, “by taking reasonable security safeguards to prevent personal data breach.” Rule 6(1) lists what those safeguards must include, at the minimum.

In practice: Rule 6(1) covers seven things: encryption, masking or virtual tokens; access control; logs, monitoring and review; backups for continued processing; keeping those logs and the personal data for one year; processor contract clauses; and technical and organisational measures. It is a minimum, not a safe harbour. Failing to take reasonable safeguards carries a penalty of up to ₹250 crore.

Source: DPDP Act, s.8(5), Schedule; DPDP Rules, r.6. Rule 6 checklist for engineers

User account

In the Rules, the online account registered by the Data Principal with the Data Fiduciary, including profiles, handles, email and mobile number.

In practice: Several Rules assume you can reach a person through their account, for example breach notices to affected people. Check that the account has a working way to contact the person.

Source: DPDP Rules, r.2.

Erasure and retention

Unless a law requires retention, a fiduciary must erase personal data when the person withdraws consent or once the specified purpose is no longer served, whichever is earlier, and make its Data Processor erase the data it was given (s.8(7)). Rule 8 also requires personal data, traffic data and processing logs to be kept for at least one year.

In practice: Erasure has to reach every copy, including processors. The Rules do not spell out backup erasure, so document when your backups expire. Keep the one-year retention duty and the erasure duty apart in your design: Rule 8(3) keeps the personal data, traffic data and processing logs for at least a year for the Seventh Schedule purposes, so hold that copy for those purposes only and erase the rest.

Source: DPDP Act, s.8(7), s.12; DPDP Rules, r.8. One-year log retention in the cloud

Cross-border transfer (section 16)

The Act lets the government restrict transfers of personal data to countries it notifies. It is a negative list, not a ban on transfers. Stricter laws still apply, for example sector rules on where payment data is stored.

In practice: As of November 2026, no country has been notified. DPDP does not require data to stay in India. Rule 15 lets the government set requirements, by order, on making transferred data available to a foreign State, and Rule 13(4) lets it keep data it specifies in India for Significant Data Fiduciaries. Hosting in India keeps you clear of any future s.16 restriction and of sector localisation rules, which is why some teams choose it.

Source: DPDP Act, s.16; DPDP Rules, r.13(4), r.15. Cross-border AI APIs

Words the DPDP Act does not define.

Search results often use these terms for DPDP. The Act and Rules do not define them, so be careful when a policy or vendor uses them as if they did.

Sensitive personal data

The DPDP Act does not define it, and it has no separate tier of sensitive data. The same duties apply to all digital personal data.

In practice: Extra duties attach to a child’s data (s.9) and to Significant Data Fiduciaries (s.10), not to a sensitive category. Sensitivity is only one factor the government weighs when it notifies an SDF. Sector regulators may set their own rules for particular data, so check those for your sector.

Source: DPDP Act, s.2, s.9, s.10. DPDP for banks and NBFCs

Anonymisation

The DPDP Act does not define it. The Act applies to data about an identifiable individual (s.2(t)).

In practice: Data that no longer identifies anyone is outside that definition, but whether your data qualifies depends on the facts, so test whether it can be linked back. Rule 6(1)(a) lists masking and virtual tokens as safeguards for personal data, so treat masked or tokenised data as personal data while you can map it back.

Source: DPDP Act, s.2(t); DPDP Rules, r.6(1)(a). PII redaction for LLMs

Data controller

This is a GDPR term. The DPDP Act does not use it. Its closest match is the Data Fiduciary, and the match is not exact.

In practice: When a vendor or contract says “controller”, ask which DPDP role they mean: Data Fiduciary or Data Processor. Fix the wording in your contracts so the roles are clear for the Indian law.

Source: DPDP Act, s.2(i), s.2(k). DPDP and GDPR for SaaS

AI governance and assurance terms.

The frameworks, reports and questionnaires that enterprise buyers and regulators ask AI companies about. Mono prepares companies and collects evidence. We do not certify anyone.

AI governance

The decisions, controls and records that show how a company’s product and team use AI: which models and vendors, what data reaches them, who can switch AI features on, how outputs are checked, and who owns the risk.

In practice: Enterprise buyers ask for this in security reviews before approving a vendor with AI features. Good governance leaves evidence you can hand over: a register of AI use, a policy, and logs.

AI governance for SaaS

Shadow AI

AI tools that staff use without approval, such as personal chat accounts, browser extensions, meeting note-takers and AI features switched on inside other software.

In practice: You cannot govern what you have not found. A shadow AI audit uses records you already have, such as sign-in grants, expense data and admin settings, and replaces risky tools with approved ones.

Shadow AI audit

ISO/IEC 42001

ISO/IEC 42001:2023, the AI management system standard. Edition 1, published 18 December 2023. A management-system standard for organisations that develop, provide or use AI.

In practice: It describes how you run AI responsibly, not how a model behaves. Annex A holds 38 reference controls, and not every control is mandatory. You choose controls from a risk assessment and justify what you leave out.

Source: ISO, iso.org/standard/42001. ISO 42001 readiness

Statement of Applicability

The document, required by clause 6.1.3 of ISO/IEC 42001, that lists the controls you need, and justifies the ones you include and the ones you leave out.

In practice: An auditor reads this first. Build it from your real AI systems and risk assessment, not from a template, or it will not survive the audit.

Source: ISO/IEC 42001, clause 6.1.3. Do you need ISO 42001?

Certification body

An independent organisation that audits you against a standard and issues the certificate. ISO does not certify. ISO/IEC 42006:2025 sets requirements for bodies that audit and certify against 42001.

In practice: Only certification bodies accredited for ISO/IEC 42001 (for example by NABCB, ANAB or UKAS) issue accredited certificates. Consultants, Mono included, provide readiness. We are not a certification body.

Source: ISO/IEC 42006:2025; NABCB 220. ISO 42001 vs SOC 2

SOC 2

An attestation report by a licensed CPA firm, under AICPA standards, on controls relevant to Security, Availability, Processing Integrity, Confidentiality or Privacy, using the Trust Services Criteria.

In practice: A SOC 2 report is a CPA’s opinion, not a certificate. Buyers read the report itself and its exceptions. AICPA has exposure drafts for updated criteria with AI considerations, open for comment until 15 January 2027.

Source: AICPA. ISO 42001 vs SOC 2

SIG questionnaire

The Shared Assessments third-party risk questionnaire. It is updated every year and lists 21 risk domains, including Artificial Intelligence. It maps to ISO/IEC 42001 and NIST AI 100-1, and it is licensed, not free.

In practice: Buyers often send it as a spreadsheet. Keep standard answers and evidence ready so the AI section does not stall a deal.

Source: Shared Assessments. SIG AI domain and CSA AI-CAIQ

CSA AI-CAIQ and AI Controls Matrix

The Cloud Security Alliance AI Controls Matrix v1.1 (22 June 2026) has 247 control objectives in 18 security domains. It includes the AI-CAIQ, a self-assessment questionnaire that is the basis for STAR for AI Level 1. It is free.

In practice: A free, structured way to answer AI security questions, mapped to ISO 42001, ISO 27001, the EU AI Act and NIST AI RMF. A self-assessment is your own statement, not an audit.

Source: Cloud Security Alliance. SIG AI domain and CSA AI-CAIQ

NIST AI RMF

The NIST AI Risk Management Framework 1.0, released 26 January 2023. It is voluntary and has four functions: Govern, Map, Measure and Manage. The Generative AI Profile, NIST AI 600-1, followed on 26 July 2024.

In practice: There is no NIST certification. You use the framework to structure your own risk work and to answer buyers who ask about it. NIST says the framework is being revised, so check the current text.

Source: NIST. NIST AI RMF for SaaS teams

RBI FREE-AI

The report of the Reserve Bank of India’s committee on a Framework for Responsible and Ethical Enablement of AI, released on 13 August 2025. It has 26 recommendations.

In practice: It is a committee report, not a binding RBI direction. Regulated lenders may still draw on it for their AI policies and vendor questions.

Source: RBI press release 2025-2026/902. RBI FREE-AI readiness

Model risk management (RBI draft)

On 24 June 2026 RBI issued draft guidance on regulatory principles for model risk management. It covers models that use AI or ML, including third-party models, for banks, NBFCs and other regulated entities.

In practice: As of November 2026, RBI has not issued final guidance, and the June 2026 draft is the latest text. Write “the draft proposes”, not “RBI requires”.

Source: RBI press release 2026-2027/528. RBI model risk and GenAI

AI security and engineering terms.

Terms from AI security reviews and from the work of building AI features and private models.

OWASP Top 10 for LLM Applications

A ranked list of the main risks in applications built on large language models. The 2025 edition has ten: Prompt Injection, Sensitive Information Disclosure, Supply Chain, Data and Model Poisoning, Improper Output Handling, Excessive Agency, System Prompt Leakage, Vector and Embedding Weaknesses, Misinformation and Unbounded Consumption.

In practice: Say “mapped to the OWASP LLM Top 10”, never “OWASP compliant”. OWASP does not certify anyone. A review checks each risk against how your features really work.

Source: OWASP GenAI Security Project. AI security review template

OWASP Top 10 for Agentic Applications

A separate OWASP list for AI agents that call tools and act on their own, published 9 December 2025. It names Agent Goal Hijack, Tool Misuse and Exploitation, Identity and Privilege Abuse, Agentic Supply Chain Vulnerabilities, Unexpected Code Execution, Memory and Context Poisoning, Insecure Inter-Agent Communication, Cascading Failures, Human-Agent Trust Exploitation and Rogue Agents.

In practice: Agents widen the attack surface because they hold credentials and take actions. Give each agent its own identity, the least access it needs, and a log of every tool call.

Source: OWASP GenAI Security Project. Questions to ask about AI agents

Prompt injection

An attack where text the model reads, in a user message, a web page, a document or a tool result, tries to change what the model does. It is LLM01 in the OWASP LLM Top 10.

In practice: You cannot filter your way out of it completely. Limit what the model can do if it is fooled: narrow tool access, no secrets in the prompt, and a human check before any action that cannot be undone.

Source: OWASP LLM Top 10, LLM01. Vibe coding security checklist

Excessive agency

Giving an AI feature more tools, permissions or freedom than its job needs. It is LLM06 in the OWASP LLM Top 10.

In practice: Start from what the feature must do and remove the rest: read-only access by default, scoped credentials, and approval steps for sending, paying or deleting.

Source: OWASP LLM Top 10, LLM06. AI architecture review questions

Model Context Protocol (MCP)

An open protocol that lets AI applications connect to tools and data through servers. In the 2026-07-28 specification, authorization is optional. HTTP transports should follow it. STDIO servers should not, and take credentials from the environment instead.

In practice: Where authorization is used, the spec requires servers to validate the token audience and not accept other tokens. For tools, the spec recommends a human in the loop who can deny invocations. Real flaws exist: CVE-2025-49596 is unauthenticated remote code execution in MCP Inspector below 0.14.1 (CVSS 4.0 score 9.4). Treat each MCP server as third-party code.

Source: modelcontextprotocol.io; CVE-2025-49596. MCP server security risks

PII redaction

Finding personal identifiers in text and masking or replacing them before the text reaches a model, a log or a third party.

In practice: For Indian identifiers, the open-source Presidio library includes recognisers for PAN, Aadhaar, vehicle registration, voter ID, passport and GSTIN. Redaction reduces leakage, so test recall on your own data. A PAN has ten characters, and an Aadhaar number has 12 digits. A masked Aadhaar shows only the last four. Aadhaar requesting entities that store full numbers must keep them in an Aadhaar Data Vault. Everyone else should still mask to the last four digits before any LLM call.

Source: Presidio; Income Tax Department; UIDAI. PII redaction for LLMs

Zero data retention (ZDR)

An arrangement under which an AI vendor does not keep your prompts and outputs after processing the request. It is usually a contract or an account setting, not a default.

In practice: Check the plan, not the brand. As of 20 November 2026, Claude Code’s zero data retention is set per organisation and is for qualified Enterprise accounts only. Cursor says that with Privacy Mode on, customer data is not used for training and it has ZDR agreements with its providers. GitHub says Copilot Business and Enterprise users are not affected by its April 2026 change to training on interaction data. Vendor terms change, so re-check before you rely on them.

Source: Vendor documentation, checked 20 November 2026. AI coding assistant data flows

LLM observability

Recording what an AI feature did on each request: the model, the tokens used, the latency, errors and, if you choose, the prompt and response.

In practice: OpenTelemetry’s GenAI conventions are an emerging standard, still marked as in development. Prompt and response content is opt-in in them. Decide deliberately whether to log content, because logs that hold prompts can hold personal data.

Source: OpenTelemetry semantic conventions for GenAI. LLM observability setup

Prompt caching

A vendor feature that stores the repeated start of a prompt so later requests reuse it at a lower price and with less delay.

In practice: Anthropic’s public pricing (checked 20 November 2026) charges more to write a cache entry and much less to read one, and OpenAI also discounts cached input. Put stable text first and changing text last, and watch the cache hit rate. Output tokens, long context and retry loops often drive cost more.

Source: Anthropic pricing page. AI spend audit

Private LLM

A language model that runs on infrastructure you control, so prompts and data do not go to a model vendor’s service.

In practice: It gives you control over data, region and access, and it adds work: hosting, updates, evaluation and monitoring. Whether it costs less than an API depends on your volume.

Own Your AI

Open-weight model

A model whose trained weights you can download and run yourself. The licence says what you may do with it, and it varies by model.

In practice: Read the licence before you build a product on it, and evaluate the model on your own tasks rather than public scores.

Open-weight model licences

About this page.

Checked 20 November 2026

Dates, figures and status lines were checked against the sources named on each entry on 20 November 2026. Standards, vendor terms and draft guidance change, so confirm the current text before you rely on it.

What Mono does and does not do

We prepare companies, build the engineering and collect the evidence. We are not a certification body, an auditor or a law firm, and nothing here is legal advice. Our partner law firm reviews the legal side of DPDP work.

Glossary FAQs.

Is this glossary legal advice?

No. It explains terms in plain words and points to the Act, the Rules and the standards. Mono is not a law firm. For advice on your situation, speak to a lawyer. On client projects our engineers do the technical work and our partner law firm handles the legal review and documentation.

Does the DPDP Act define sensitive personal data?

No. The Act does not set up a separate sensitive category, and it does not define anonymisation or data controller either. It sets extra duties for children’s data and for Significant Data Fiduciaries.

Do I have to use a Consent Manager?

No. Consent Managers are optional for fiduciaries. You can run your own consent flow, as long as you can show notice and consent. The Board’s Consent Manager registration portal is live at dpbi.gov.in.

Am I a Significant Data Fiduciary?

Only if the Central Government notifies you under section 10 of the Act. As of November 2026, none have been notified.

When do the main DPDP duties start?

Notice, consent, rights, Rule 6 safeguards, breach duties and retention and erasure start on 13 May 2027. Some provisions already apply, so check the date for the duty you care about.

Does Mono certify companies for ISO 42001, SOC 2 or DPDP?

No. We are not a certification body or an auditor. Certificates come from accredited certification bodies, and SOC 2 reports from licensed CPA firms. There is no government DPDP certificate. We prepare you, build the engineering and collect the evidence.

Not sure which terms apply to your systems?