Confidential Computing and TEEs in Software Development: A Secure Room for AI

Published on 4 November 2026

Artificial intelligence has already become an essential companion in software development. We ask it to understand thousands of lines of code, analyse repositories, interpret incidents, and help modernise applications. To do so, we grant AI access to contextual information that may contain sensitive assets, such as source code, credentials, or business data. While this information can be encrypted both at rest and in transit, it must be decrypted at some point for the AI to use it. This is where Confidential Computing comes in: a set of technologies designed to protect data not only when stored or transmitted, but also while it is being processed. Its core component is the Trusted Execution Environment (TEE).

What is a TEE?

A TEE can be understood as a protected execution environment. Information enters, is processed within it, and the result leaves, but sensitive data remains protected throughout its use. We can think of it as a "secure room" within the processor’s memory, isolated from the rest of the system. Neither the operating system nor even a cloud provider’s administrator can access, read, or alter the information inside.

In addition, when AI is expected to work with sensitive information, TEEs provide attestation mechanisms. These allow us to verify that the environment, software and configuration are exactly as expected before any information is shared with it.

What Role Does Software Development Play?

As you may have guessed, Confidential Computing impacts multiple disciplines. Infrastructure and Cloud teams provide TEE-enabled servers, virtual machines, containers and services. Cybersecurity teams manage identities, encryption keys, secrets, policies and controls. DevSecOps incorporates these requirements into deployment, testing, observability and operations.

For software development teams, however, TEEs are fundamentally an architectural and design concern. We must decide which parts of an application, which data sets, processes or components require additional protection so that only those specific elements are isolated. Running an entire application within a TEE would rarely make sense. Instead, we need to identify and separate sensitive components, properly manage secrets and credentials, and integrate attestation mechanisms where appropriate.

For example, in an AI-assisted development platform, it may be appropriate to protect the component that retrieves code from repositories, the agent that analyses incidents, the engine that processes traces, or the service responsible for preparing context before an AI model performs inference.

A Practical Use Case

Imagine that an organisation introduces AI to support the processing of administrative cases. The AI reads documents, accesses financial information, consults personal data, and applies internal rules to generate a response or recommend next steps. While the AI does not need permanent access to this information, it does need to process it in order to provide useful results.

If this part of the workflow runs within a TEE, documents and data are decrypted only inside the protected environment. Before any information is provided, the system verifies that the authorised AI agent is running with the correct software version and configuration. Once the analysis has been completed, only the necessary output leaves the protected environment to continue the process.

This highlights an important software design decision: identifying which data requires special handling so that the appropriate infrastructure-level protections can then be applied.

In Summary

A TEE is a protected execution environment that enables AI systems to work with sensitive information while keeping that information isolated during processing. It extends data protection to the moment when an application or AI system is actively using the information, providing an additional layer of security when handling particularly sensitive data.

In short, it represents another important step towards strengthening confidentiality when designing and developing AI-powered applications.