Security teams cannot secure what they don’t know is there, says Richard Frost, Head of Technology Solutions and Consulting at Armata Cyber Security.
Shadow AI has become one of the most prevalent forms of shadow IT in 2026, with unsanctioned AI tools adopted by employees faster than IT departments can discover or govern them. It is creating a data exposure problem that has become larger in scale than the shadow IT problem that spawned it. The Netskope Cloud and Threat Report that analysed cloud security from October 2024 to October 2025 found that 47% of generative AI users were accessing these tools within the enterprise environment through personal and unmanaged accounts that didn’t align with any of the company’s security controls.
The cost of this is also on the increase. The IBM Cost of a Data Breach Report has found that companies with a high shadow AI percentage experience an increase in breach costs. The same data also found that the percentage of security incidents using shadow AI was 43%. These are not the only reports highlighting the seriousness of the trend. The Verizon 2026 Data Breach Investigations Report – employees using AI tools on corporate devices increased to 45% in a year; and the ISACA 2026 study found that companies are still operating without an active AI policy.
Shadow IT isn’t a new phenomenon. It has been around since the first employee thought it was a good idea to access their own technology on their own devices within the corporate network. However, it is changing shape, moving from a relatively known threat parameter that had controls in place to mitigate employee error to one that is pervasive and pernicious. Only two out of every 10 employees are not using AI tools that haven’t been approved by the business. The other eight are openly using the AI tools they prefer despite the risk. The speed of AI adoption and integration is moving faster than governance and security can keep up.
The problem is that employees are opting for tools that are convenient. They may not like or know how to use the corporate-approved systems, so instead they revert to the tools they have grown used to. And almost everyone is guilty of this – quickly jumping onto ChatGPT instead of using Copilot because the former is faster or has more features – but the real problem comes down to what information they are putting into these AI systems. Many people don’t realise that the AI they bring to work doesn’t operate within a secure, walled garden and that the data they are putting inside the AI isn’t necessarily secured.
The risks here are twofold. On one hand, there is the risk of losing control of the data shared on an unsanctioned AI. You don’t know where it is stored, how it is processed or who can access or leak it. If an employee uploads sensitive information like customer contacts, records or financial models to a public AI service, then the company has no contractual control over its retention or the downstream sharing of that data. While the Information Regulator is still determining AI policy, there is a risk down the line that this sharing of information with an unregulated AI will constitute a breach of POPIA.
Addressing shadow AI shouldn’t come down to banning people from using it – that will only make them use it more often, and more secretively. There are several ways to approach it. The first is to make the unsanctioned AI into a sanctioned tool, which can be an expensive investment; the second is to make the sanctioned tool even better than the unsanctioned one. But this approach only works alongside an actual policy, which many companies still don’t have. The policy needs to be enacted and enforced, and it needs to clearly outline what constitutes the hygienic use of AI across prompting, sharing, use cases and more. When you can pair a licensed and governed version of the tools people already want to use with a written AI policy, the incentive to look for a workaround mostly disappears.
Companies cannot secure what they don’t know isn’t there, and banning systems has never made them disappear, so getting ahead of the shadow AI trend means moving fast to govern them, license them and write the policy while recognising the risk and putting security systems in place that can catch unsanctioned use at speed. This is where Zero Trust-oriented security providers add value because rather than trying to block every unsanctioned tool, they focus on discovering shadow AI traffic, applying identity-aware access controls to sanctioned alternatives, and inspecting data in motion before it reaches an ungoverned model. For a South African enterprise operating under POPIA, that capability directly maps to legal obligations: demonstrating a lawful basis for processing, maintaining an auditable trail of where personal information travels, and closing the operator-contract gap that consumer AI tools leave wide open.
