The Agentic Frontier: Securing AI Agents in Indian Financial Services

The security of AI agents in India's financial services sector is paramount. AI agents can perform actions that are distinct from those of AI chatbots. These actions can include making payments, altering risk profiles, or rejecting insurance claims based on the information they process. Boards, CISOs, and risk teams must effectively govern these AI agents to mitigate potential risks.

AI agents can pose "agentic risk" when they have access to data, utilize tools, delegate tasks, or change systems within a workflow. An error in a single step of an AI workflow can propagate throughout the entire process, leading to incorrect customer decisions. Therefore, the security boundary for AI agents should ideally be established outside the model itself.

Indian financial institutions are advised to review their AI inventories, audit coverage, agent permissions, and incident-response plans. The CERT-In has published AI Security Auditing Guidelines on November 4, 2026, which provide a framework for assessing AI systems. Institutions should verify the relevance of these guidelines to their specific operations before considering them as mandatory requirements.

For banks and non-banking financial companies (NBFCs), it is crucial to assess the agentic underwriting, collections, and third-party AI against Reserve Bank of India (RBI) directions and fair-practices obligations. Vendor or autonomous agents do not absolve the regulated institution of its own accountability. Securities-market participants must also consider SEBI's Cybersecurity and Cyber Resilience Framework for evaluating covered systems, monitoring, and incident response.

Insurers should evaluate automated underwriting and claims procedures against the guidelines set by the Insurance Regulatory and Development Authority of India (IRDAI). In the event of a rejected claim, the evidence, policy terms, and decision owner should be traceable to ensure accountability. Personal data should be mapped throughout the AI workflow, including prompts, agent memory, retrieval indexes, logs, and vendors, to ensure compliance with the Data Protection and Privacy Act (DPDP) and other sectoral regulations.

Five key controls should be implemented to enhance the security of AI agents:

1. Assign each agent a distinct, non-human identity.

2. Grant access to specific tools and records for the task, rather than sharing broad service credentials.

3. Separate recommendations from execution; an independent policy gate should evaluate the action, customer, amount, and authority before the API executes it.

4. Define approval thresholds for high-impact actions, such as payments, changes to risk profiles, claims exceptions, and production modifications. Reviewers should have access to the evidence, not just an approval button.

5. Establish a robust stop mechanism capable of blocking tool calls, revoking credentials, suspending queued work, and testing the impact on ongoing transactions. An independent audit trail should be maintained, linking agent identity, source data, tool calls, policy checks, approvals, and outcomes throughout the workflow.

A practical test for boards involves selecting one agent with access to a financial system and asking three questions: Who authorized it, and what can it do right now? Can we demonstrate why its last material action was permitted? Can we prevent it from initiating the next action before it affects customers or payment systems? If the answers are unclear, the institution may have identity and control issues with its AI agents, rather than solely a model problem.