The more data you feed your AI ecosystem, the more exposed your assets become, warns Chris Badenhorst, VP of Microsoft Azure and AI at Braintree.
Most companies know where their contracts, customer records and essential data are kept, yet few can say how much of this information has passed through an AI tool.
Summarised reports, draft strategies, spreadsheets – these are fed into the AI, which has slowly evolved to become a concentrated storehouse of all your company’s most sensitive thinking. In many organisations, especially small to medium enterprises (SMEs), AI was switched on first and the question of protecting the data it contained came later. As the Microsoft Data Security Index found, 32% of the incidents experienced by the companies surveyed were as a result of Generative AI.
The retrofitting of AI into the business is a global problem, not just a South African one. Most companies have had to stop and think about their level of exposure as AI seeps into every part of their business. Morgan Stanley’s 2026 global research found that around 90% of surveyed companies believe that data is the most common source of risk and 56% believe this comes from their use of AI, yet only 41% had a complete inventory of the AI models in use with assigned ownership. Data oversharing, leaks and vulnerabilities are a very real concern for the modern business, regardless of size.
And the routes taken by cybercriminals into the business to access its AI systems are tried and trusted. Phishing for usernames and passwords still works with Microsoft’s 2026 Digital Defence Report finding that valid accounts make up 20% of the initial access into the business. AI, says the report, ‘does not erase the security lessons of the past, it is making familiar weaknesses easier to exploit’. A stolen login to an AI account opens the door to whatever was uploaded into the account, including the history of everything that was asked of it.
While the employee has always been one of the weakest links in the security chain, the ways in which people use AI have made it even riskier. If someone takes a document off OneDrive, for example, and drops it into their personally preferred AI tool instead of the one mandated by the company, they are effectively bypassing all the security controls put in place by the business. Shadow AI (the use of AI tools without IT approval) is widespread, particularly in organisations where the sanctioned AI tools are restricted or less capable than the latest frontier models. When employees feel limited, they often turn to tools of their own choosing. Unvetted tools like these pose a real security risk: they may contain malware or spyware or hidden code that allows them to carry out unauthorised actions once installed on a user’s machine.
Every one of these shadow AI tools is another potential open door into the business, one that is made even riskier as documents and data are thrown into them.
This may seem innocuous at first, but imagine if your company was in the process of developing a solution that could put your business on the map and then your team is working through problems and solutions using a public AI model. Depending on the service and the settings, the data fed to the model can be used to train it and you lose complete control over where that thinking travels next. Researchers have demonstrated extraction of memorised training material from language models, including ChatGPT, using specially designed attacks.
Regulation adds the final layer of concern, especially as many companies are still working out how compliance applies when AI is involved. In South Africa, POPIA already requires that personal information be secured with appropriate and reasonable technical and organisational measures, so meeting that requires visibility and robust security. Identity, access and data controls need to hold across the entire company and this means multi-factor authentication on every account that touches AI and conditional access that judges each sign-in on its context.
Right now, leadership teams have two questions to answer as they wrestle with AI. The first is who decides what data employees can share with AI and how that is enforced, especially when there are deadlines. The second is ownership – you may legally own every document your employees upload but if they sit in an account you don’t control, getting them back after a lockout depends on the login and the provider. Securing your organisation requires that AI be governed, managed and monitored with controls that are stringent enough to protect your business and your data but don’t inhibit the potential of AI.
And that balance is still in progress.
