Enterprise security has spent two decades getting very good at protecting data in two of its three states. Data at rest is encrypted on disk. Data in transit is encrypted across the network. The third state — data in use, the moment it is decrypted into memory and actually processed — has quietly been left in the open, because for most workloads the exposure was brief and hard to exploit. AI ends that comfort. A model does not glance at sensitive data; it reads all of it, repeatedly, holding payroll, health records, financial data and enterprise IP in live memory for as long as it trains or infers. At L&T Vyoma, our position is that traditional cybersecurity is no longer sufficient on its own for AI, and the missing piece has a name: confidential computing. In India, where that same data carries sovereignty obligations under the DPDP Act, closing the data-in-use gap is both a security decision and a compliance one.
AI Broke the Security Model It Inherited
Traditional cybersecurity was designed around applications and infrastructure — firewalls at the perimeter, endpoint protection on the device, monitoring in the SOC, patching for known vulnerabilities. That model assumed data spent most of its life stored or moving and only briefly in use. AI inverts the assumption. Training and inference exist precisely to process data, at volume and continuously, and the data they process is overwhelmingly the sensitive kind: customer information, financial records, healthcare data, government and regulated datasets, and the enterprise's own intellectual property. Every stage of the AI lifecycle, from training through inference to the prompts users type, becomes a new point at which sensitive data sits decrypted and readable. Security built for the application era does not cover the AI-era risk, because it was never meant to protect data while a model is actively computing on it.
The Missing Layer: Protecting Data While It's in Use
The exposure is already measurable. Harmonic Security's GenAI Data Exposure study, which analysed one million generative-AI prompts and twenty thousand uploaded files in the second quarter of 2025, found that nearly 22% of uploaded files and 4.37% of prompts contained sensitive content, most often source code, alongside customer profiles, financial data and employee personal information. That data does not stay at the keyboard; it flows into models running on infrastructure the enterprise usually does not control. Encryption at rest and in transit does nothing for it, because the exposure happens in between, while the data is decrypted in memory and being processed. Hardware security modules protect the keys, not the data those keys decrypt once it is in use. Data-in-use protection is the layer that has been missing, and AI is what made its absence expensive.
Why Confidential Computing Closes the Gap
Confidential computing is the hardware answer to the data-in-use problem. It processes sensitive data inside a hardware-based trusted execution environment — a protected enclave on the chip itself, so the data and the code stay protected even while they are being computed on, verified by cryptographic attestation before anything is decrypted into them. For AI specifically, NVIDIA Confidential Computing extends this to the GPU: on NVIDIA Hopper and Blackwell GPUs, the class of hardware that runs the L&T Vyoma AI Factory, it protects GPU execution, memory and register states, keeping model weights, training data and inference prompts isolated at runtime. NVIDIA describes it as closing the last gap in enterprise data security, and says a model can be moved into a protected enclave without code changes, at up to 98% of unsecured performance. That last figure is NVIDIA's own, but it retires the old objection that confidential computing was too slow to use in production. The market has moved on the same logic: ABI Research projects confidential computing revenue will reach USD 160 billion by 2032, a 44% compound annual growth rate, with GPU-based confidential computing for AI the force reigniting demand.
Security-as-a-Service Is Evolving — and Compliance With It
Confidential computing does not arrive alone. The security services enterprises buy are broadening from the traditional set of firewall management, endpoint security, SOC monitoring and vulnerability management into an AI-era set: AI security, data security, encryption and key management, confidential computing, compliance management and multi-cloud security governance. The driver is partly regulatory. Compliance is shifting from a periodic event to a continuous obligation: the DPDP Act, the RBI cyber security framework, SEBI's CSCRF and CERT-In's directions increasingly expect controls that are demonstrable at any moment, rather than only at audit time. A posture that can prove cryptographically, through attestation, that sensitive data was never exposed even during processing is a far stronger answer to a regulator than a policy document asserting the same thing.
In India, Data-in-Use Protection Is a Sovereignty Question
For an Indian enterprise, where AI processes sensitive data is inseparable from where that data is allowed to live. Data residency and data sovereignty are now security priorities as much as legal ones, and the DPDP Act ties the two together: personal data processed by an AI model is still personal data, subject to the same consent, purpose and cross-border rules. Confidential computing strengthens the sovereignty case in a way geography alone cannot: it gives a hardware-backed guarantee that not even the infrastructure operator, or a privileged insider, can read the data while it is in use. Run that on sovereign, in-country infrastructure and the two controls compound. This is the combination L&T Vyoma is built to provide: confidential AI on NVIDIA Hopper and Blackwell GPUs, inside L&T-operated data centers on Indian soil, so an enterprise can train or infer on its most sensitive data with the data-in-use gap closed and the whole workload staying in the country, provably, under DPDP.
|
State of data |
Where it is exposed |
What protects it |
|
At rest |
Stored on disk or in a database |
Encryption at rest (long solved) |
|
In transit |
Moving across the network |
TLS / encryption in transit (long solved) |
|
In use |
Decrypted in memory while an AI model processes it |
Confidential computing: a hardware trusted execution environment |
Security for AI Has to Be Built In, Not Bolted On
The pattern across every serious AI security discussion is the same: the protections that hold are the ones designed into the infrastructure, not added afterward. Zero-trust architecture, continuous compliance and confidential computing are becoming the baseline for running AI at scale on sensitive data, and data-in-use protection is the piece that turns “we secured the perimeter” into “we can prove the data was never exposed.” Traditional cybersecurity still matters; it is necessary, simply no longer sufficient for AI. The enterprises that adopt AI safely will be the ones that treat confidential computing as the default for sensitive workloads rather than a special case. Talk to our team about running confidential AI on sovereign, in-country infrastructure, or explore the L&T Vyoma AI Factory and our security services.
Sources: sensitive data entering generative-AI tools — Harmonic Security, GenAI Data Exposure report, Q2 2025 (~22% of uploaded files and 4.37% of prompts contained sensitive content); confidential computing market size and growth — ABI Research, 2025 (USD 160 billion by 2032, 44% CAGR; GPU-based confidential computing reigniting demand); data-in-use protection on NVIDIA Hopper and Blackwell GPUs — NVIDIA (Confidential Computing solution and glossary); DPDP Act, 2023, with RBI cyber security framework, SEBI CSCRF and CERT-In directions for the continuous-compliance context.