As Indian enterprises move AI from pilots into production, Cybersecurity Awareness Month is a reminder that AI can only be scaled as fast as it can be secured and observed.
Indian enterprises are moving AI out of the pilot stage. Banks are using it in customer service and fraud detection, manufacturers in quality control, and hospitals in diagnostics and patient records. As these systems begin to handle real data and real decisions, a basic question follows: can the organisation see what its AI is doing, and can it stop that AI from being misused?
For most, the answer is not yet. Cisco’s 2025 Cybersecurity Readiness Index found that 95% of organisations in India had experienced an AI-related security incident in the previous year. In the same study, 45% said they were not confident they could detect unapproved AI deployments inside their own business.
An organisation cannot protect, govern or trust a system it cannot see. That makes visibility the first requirement of any AI programme, not a later addition.
Two kinds of AI risk
Conventional software behaves predictably. Given the same input, it returns the same output. AI models do not work that way. Their responses vary, and they can change as models are updated or retrained. This creates two distinct categories of risk that tend to reinforce each other.
The first is safety risk: the model itself behaves in ways the business did not intend. It may produce inaccurate or biased answers, reveal sensitive data in its responses, or act outside the limits set for it. The second is security risk: an outside party deliberately manipulates the model. Common techniques include prompt injection, attempts to bypass a model’s restrictions, poisoning of training data, and theft of the model or the data behind it.
Most existing security controls were built to inspect network traffic, files and user behaviour. They were not designed to examine a prompt, judge a model’s response, or follow an AI agent as it calls other systems. That gap is where much of the new risk sits.
The cost of losing visibility
The financial impact is already measurable. IBM’s Cost of a Data Breach Report 2026 puts the average cost of a breach in India at ₹25.5 crore, a record and 15.9% higher than in 2025. Shadow AI, meaning AI tools used without the knowledge or approval of IT, added an average of ₹1.79 crore to the cost of a breach where it was present. It ranked among the three largest factors that raised breach costs in India.
The regulatory exposure is also growing. Under the Digital Personal Data Protection Rules, core obligations take full effect on 13 May 2027. These include reasonable security safeguards and the reporting of personal data breaches. An AI system that processes personal data is subject to the same obligations as any other system, and an organisation that cannot see how its AI handles that data will find them hard to meet.
The larger cost is trust. Customers, regulators and partners extend confidence to a business on the basis of how it handles their data. When an AI system leaks information or behaves in a way the business cannot explain, that confidence is lost quickly and recovered slowly.
Observability and security as one discipline
Organisations that scale AI safely tend to treat observability and security as a single discipline, applied across the full life of an AI application.
- Discover. Build an inventory of every AI application, model and agent in use, including those adopted by business teams without approval.
- Validate. Test models before deployment for vulnerabilities, unsafe behaviour and susceptibility to manipulation, and repeat the testing whenever a model changes.
- Protect. Apply guardrails while the application is running, checking both the inputs sent to a model and the responses it returns.
- Observe. Monitor the full stack, from network and infrastructure to applications, AI workloads and the experience of the end user. Problems can then be traced to their source in minutes rather than days.
Published frameworks give this work a common structure. These include the OWASP Top 10 for Large Language Model Applications, the NIST AI Risk Management Framework and MITRE ATLAS. Aligning to them gives boards, auditors and regulators a shared reference point.
The technology to support this is now available. Cisco AI Defense, for example, is designed to discover AI applications in use, validate models and protect them at runtime. Splunk Observability provides real-time visibility across infrastructure, applications and user experience. Used together, they give security and operations teams a common view of AI systems in production.
A shared responsibility
No single organisation can secure AI on its own. Technology vendors build the controls. System integrators and solution partners design, deploy and manage them for customers. Most Indian businesses, from mid-sized manufacturers to regional banks and hospitals, depend on that partner ecosystem rather than on large in-house teams.
This is where distribution plays a part beyond supply. Technology distributors work with partners across India to design networking and security architectures across portfolios, providing end-to-end customer support, and run certification and training programmes. This empowers a provider in a tier-two city to deliver the same standard of AI security and observability as one in a metro.
Cybersecurity Awareness Month is a useful point for leadership teams to ask three questions:
- Do we know every AI application, model and agent in use across the business?
- Can we see what goes into those systems and what comes out of them?
- If a model behaves unexpectedly or is manipulated, how quickly would we know, and who is accountable?
Organisations that can answer these clearly are in a position to scale AI with confidence. Those that cannot should address the gaps before they expand AI further. Either way, the trust placed in AI depends on the security and visibility built into it from the start.
Authored by Hetal Shah, Head – Technology Solutions Group (TSG), Redington India
