Quick answer
AI agents that call tools — CRMs, email, databases, payments — process personal data on every step, and each action must stay within the purpose and lawful basis the Data Fiduciary has. Under the DPDP Rules, build least-privilege tool credentials (Rule 6(1)(b)), action-level logs retained at least one year (Rules 6(1)(c), 8(3)), human approval for disclosures, and erasure that reaches agent memory.
- An agent acts for the Data Fiduciary; every tool call it makes is processing you are responsible for.
- Scope each tool to a purpose, with least-privilege, short-lived credentials.
- Require human approval before an agent discloses personal data externally or takes irreversible actions.
- Log every tool call with inputs, outputs, purpose and approver for at least one year.
- Prompt injection that makes an agent exfiltrate data is a personal data breach under Rule 7.
- Agent memory is a personal data store and needs retention limits and erasure.
Who this is for: Teams building agentic workflows, Platform and security engineers, Operations leaders automating back-office work, CISOs, Data Protection Officers
Why AI agents raise different DPDP issues
A chatbot answers. An agent acts: it reads a customer record, drafts and sends an email, updates a ticket, books a refund, queries a database, browses a site or calls another agent. Each of those steps is processing of personal data — retrieval, use, disclosure, alteration — and the Data Fiduciary that deployed the agent is responsible for all of it under section 8(1), including steps carried out through vendor tools.
What changes compared with a chatbot:
- Breadth of access. Agents are often given broad API keys "so they can be helpful", which collapses purpose limitation.
- Autonomy. The model chooses which data to fetch and where to send it, so a confused or manipulated agent can disclose data no human intended.
- Chaining. Data fetched by one tool flows into another tool's input — for example, from a CRM into an outbound email or a third-party search API.
- Memory. Persistent agent memory becomes a new personal data store, often outside existing retention and erasure processes.
- Attack surface. Indirect prompt injection — instructions hidden in a web page, email or document — can steer an agent to exfiltrate data.
The DPDP Act has no agent-specific rules, but its general duties apply directly: purpose and lawful basis, reasonable security safeguards (s.8(5) and Rule 6), breach intimation (Rule 7), erasure (s.8(7)) and processor contracts (s.8(2)).
Scope tools to purposes, not to the agent
Design tool access as you would design access for a new employee who will follow any instruction literally.
| Control | Implementation |
|---|---|
| Purpose-bound toolsets | Each agent has a declared purpose_id; only tools approved for that purpose are registered |
| Least-privilege credentials | Tool calls run under a service identity scoped to the minimum records and operations, or under the end user's delegated token |
| Short-lived tokens | Issue per-task credentials that expire; never embed long-lived admin keys in agent config |
| Field-level filtering | Tool wrappers return only the fields the task needs, masking others (Rule 6(1)(a)) |
| Read/write separation | Separate read tools from write and send tools, with different approval rules |
| Egress allow-lists | Agents may call only approved domains and APIs; block arbitrary URLs for data-bearing requests |
| Rate and volume limits | Cap records fetched per task; alert on bulk reads about many different people |
Rule 6(1)(b) requires appropriate measures to control access to computer resources used by the Data Fiduciary or its processor. For agents, the tool gateway is where that control lives. Put every tool behind it, rather than letting the agent framework hold credentials directly.
Human approval and irreversible actions
Not every step needs a human, but some must. Use a risk-tiered approval policy:
- Auto-approve: reading data the end user is already entitled to see, within the task's scope.
- Approve before execution: sending personal data outside the organisation (email, messaging, third-party APIs), changing records that affect a person's account, money movement, deletion.
- Never automate: decisions with legal or similarly significant effect on a person without a human decision-maker, where accuracy duties under section 8(3) apply — for example, rejecting a claim or a job application.
Show the approver exactly what will be sent and to whom, not a model summary. Record the approver's identity and the payload hash in the action log.
Consent and purpose checks belong in the same gate. Before a "send marketing follow-up" tool runs, check the recipient's consent status for that purpose and whether she has withdrawn consent. Under section 6 and Rule 3(c)(i), withdrawal must be as easy as giving consent, and processing must stop within a reasonable time after it — an agent that keeps emailing a withdrawn customer is a compliance failure, not a bug.
For children's accounts, block any tool that would enable tracking, behavioural monitoring or targeted advertising (section 9(3)).
Action logs: the evidence trail
Rule 6(1)(c) requires visibility on the accessing of personal data through appropriate logs, monitoring and review, for detecting unauthorised access, investigating it and preventing recurrence. Rule 8(3) requires processing logs to be retained for at least one year. For agents, log at the level of each step:
{
"run_id": "...", "step": 7, "agent_id": "refund-agent",
"purpose_id": "order-support", "on_behalf_of": "user:pseud-8812",
"tool": "crm.get_customer", "args_redacted": {"customer_id": "tok_91f"},
"records_returned": 1, "fields": ["name","order_history"],
"approval": null, "egress_domain": null,
"model": "provider/model-version", "ts": "2027-05-20T10:14:03+05:30"
}
Guidelines:
- Log identifiers and field names, not raw values, in the main log; keep raw payloads, if needed, in a restricted tier.
- Synchronise clocks — CERT-In Directions require clock sync to NIC or NPL NTP servers or traceable sources, and keep ICT logs for a rolling 180 days within India.
- Review logs: alert on new egress destinations, unusual volumes, and tool sequences that read personal data and then send externally.
This trail answers the Board's likely questions after an incident: what did the agent access, for whom, on what authority, and where did it send it.
Prompt injection and breach response
Indirect prompt injection is the agent-era version of a phishing attack. An agent reads a support email that contains hidden instructions to forward the customer list to an outside address, and complies. Under the DPDP Act, that is a personal data breach — unauthorised processing and disclosure compromising confidentiality.
Prevent
- Treat all retrieved content (emails, web pages, documents, tool outputs) as untrusted data, never as instructions.
- Keep tool permissions narrow so a manipulated agent cannot do much.
- Require approval for external sends, as above.
- Red-team agents with injection test suites before release and after changes; Rule 6(1)(g) calls for technical and organisational measures to ensure safeguards work.
Respond
Rule 7 applies in full:
- Intimate each affected Data Principal without delay, through her user account or registered contact, with the nature, extent and timing, likely consequences, mitigation, steps she can take and a contact.
- Intimate the Board without delay with a description, then within 72 hours give detailed information: facts, mitigation, findings on who caused it, remedial measures and a report on intimations.
- Report to CERT-In within 6 hours if it is a specified cyber incident.
See the breach response playbook.
Agent memory, multi-agent systems and vendors
Memory
Long-term memory — user preferences, past tasks, extracted facts — is a personal data store. Give it a schema (data_principal_id, purpose_id, created_at, source_run_id), a retention limit tied to the purpose, user-visible controls where it is customer-facing, and inclusion in erasure workflows under sections 8(7) and 12.
Multi-agent and cross-organisation calls
When your agent calls another organisation's agent or API, you are disclosing personal data to a third party. Decide whether that party is your Data Processor (contract required, section 8(2)) or a separate Data Fiduciary (a lawful basis for sharing required, and the identity disclosed in access responses under section 11).
Vendors
Agent frameworks, hosted orchestration, browser-automation services, model APIs and observability tools are processors. Contract for security, no training on your data, retention, deletion, breach notice and region (Rule 6(1)(f)). For cross-border calls, the DPDP Act permits transfers except to notified restricted countries (s.16), subject to Rule 15 and sector rules.
For SEBI-regulated firms, remember the regulated person remains solely responsible for AI tool outputs and data security regardless of vendor. See SEBI AI/ML rules.
Further reading on suhasbhairav.com
Before 13 May 2027
Compliance checklist
Declare a purpose for every agent
Register each agent with a purpose ID and lawful basis, and attach only tools approved for that purpose.
Sections 4, 6, 7
Put all tools behind a gateway
Route tool calls through a gateway that enforces credentials, field filtering, egress allow-lists and volume limits.
Rule 6(1)(b)
Use least-privilege, short-lived credentials
Issue per-task or delegated user tokens instead of long-lived admin keys.
Rule 6(1)(b)
Gate external disclosures with human approval
Require approval with the exact payload before sending personal data outside the organisation.
Section 8(5)
Check consent before outreach tools run
Block marketing or follow-up actions for people who withdrew consent or are children.
Section 6; Section 9(3)
Log every agent step for at least one year
Record tool, arguments (redacted), records returned, approvals and egress for each step.
Rule 6(1)(c); Rule 8(3)
Red-team for prompt injection
Test agents against injected instructions in emails, documents and web pages before each release.
Rule 6(1)(g)
Bring agent memory into retention and erasure
Schema-tag memory entries, set retention and delete on erasure requests.
Sections 8(7), 12
Add agent incidents to breach runbooks
Define detection, Rule 7 intimations and CERT-In 6-hour reporting for agent-caused leaks.
Rule 7
Key dates
DPDP Rules notified
Security, logging, breach and erasure duties apply to agentic systems
Answers
Frequently asked questions
Who is responsible when an AI agent misuses personal data?
Who is responsible when an AI agent misuses personal data?
The Data Fiduciary that deployed the agent. Under section 8(1) of the DPDP Act it is responsible for processing done by it or on its behalf, including through vendor tools. The agent's autonomy does not shift liability; design permissions, approvals and logs accordingly.
Can an AI agent send customer data to third-party APIs?
Can an AI agent send customer data to third-party APIs?
Only if the disclosure fits the purpose and lawful basis, the recipient is under a processor contract or has its own lawful basis, and appropriate safeguards are in place. Require human approval for external sends of personal data and restrict agents to an allow-list of destinations.
What should an AI agent log under the DPDP Rules?
What should an AI agent log under the DPDP Rules?
Log each step: agent and run ID, purpose, user on whose behalf it acted, tool called, redacted arguments, records and fields returned, approvals and egress destination. Rule 6(1)(c) requires logs for detecting unauthorised access, and Rule 8(3) requires processing logs for at least one year.
Is a prompt injection data leak reportable in India?
Is a prompt injection data leak reportable in India?
Yes, if it causes unauthorised disclosure of personal data. It is a personal data breach under the DPDP Act, so Rule 7 requires intimation to affected Data Principals and the Board, with detailed information within 72 hours. CERT-In's 6-hour reporting may also apply.
Does agent memory count as personal data?
Does agent memory count as personal data?
Yes, where it stores information about identifiable people, such as preferences or facts from past tasks. Treat memory as a personal data store with a purpose, retention limit, access control and inclusion in erasure workflows under sections 8(7) and 12.
Do AI agents need human oversight under the DPDP Act?
Do AI agents need human oversight under the DPDP Act?
The Act has no explicit human-oversight clause, but section 8(3) requires accurate and complete data where it is used for decisions affecting a person, and Rule 6 requires effective safeguards. Human approval for disclosures and significant decisions is the practical way to meet both.
Legal references and sources
- Digital Personal Data Protection Act, 2023 — Sections 8(1), 8(5) (official text)
- Digital Personal Data Protection Rules, 2025 — Rule 6(1)(b), (c) (official text)
- Digital Personal Data Protection Rules, 2025 — Rule 7 (official text)
- Digital Personal Data Protection Rules, 2025 — 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 dataDPDP Act for AI Systems: Implementation Guide for Indian CompaniesHow the DPDP Act and Rules apply to LLM apps, RAG, agents and ML pipelines in India, with the architecture controls to build before 13 May 2027.
- AI & personal dataDPDP Compliance for RAG Chatbots and Knowledge AssistantsRAG chatbots copy personal data into chunks, embeddings and chat logs. This guide maps DPDP duties to ingestion, retrieval, generation and deletion.
- 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.
- 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.
- AI & personal dataAI Vendor Due Diligence Checklist Under the DPDP ActA practical due diligence checklist for LLM APIs, vector databases, observability tools and AI SaaS vendors, mapped to DPDP section 8 and Rules 6, 7 and 8(3).
- India AI & sector rulesCERT-In Rules for AI Systems: Incident Reporting & LogsHow CERT-In's April 2022 directions on 6-hour incident reporting, 180-day logs in India and clock sync apply to LLM gateways, model APIs and AI agents, and how they meet DPDP.