As employees and AI agents gain access to sensitive information, businesses are struggling to track where data goes, who can reach it and what automated systems may do with it.
As businesses race to embed artificial intelligence into daily work, many are discovering that their data is moving faster than their controls. Employees now use AI to condense reports, interrogate spreadsheets and draft strategy material, but that convenience comes with a quieter problem: sensitive information can end up in systems the organisation does not fully oversee. The result is not simply a broader attack surface, but a loss of visibility over where company information travels once it leaves internal platforms.
Microsoft’s Data Security Index suggests the scale of the issue is already material, with 32% of incidents among surveyed companies linked to generative AI. Morgan Stanley’s 2026 global research points in the same direction: about 90% of surveyed firms said data was their most common source of risk, 56% tied that risk to AI, and only 41% said they had a complete inventory of the AI models in use with ownership assigned. That gap matters because a company cannot protect what it cannot identify, and it cannot assess exposure if it does not know which tools staff are using or what information those tools receive.
Traditional cyber threats remain part of the picture, and AI has not made them disappear. Microsoft’s 2026 Digital Defense Report found that valid accounts accounted for 20% of initial access to businesses, underscoring how often attackers still rely on stolen credentials rather than sophisticated exploits. In practice, that means a compromised AI login can expose prompt histories, uploaded documents and other material shared during routine work. Security teams increasingly have to treat access to AI services with the same seriousness as access to email, file storage and finance systems.
The risks are compounded by shadow AI, the unofficial use of tools outside IT approval. An employee who downloads a confidential file from OneDrive and uploads it to a consumer AI service may bypass security controls, retention rules and monitoring systems without intending to do so. The issue is not limited to misuse of legitimate platforms. Unapproved apps and browser extensions can also introduce malware, spyware or other code that creates a direct path onto a device. The UK’s Information Commissioner’s Office recently reprimanded the Metropolitan Police Service after officers used an unapproved platform to share sensitive data without proper assessments or approvals, a reminder that shadow IT and shadow AI can have regulatory as well as security consequences.
The challenge is now visible in sectors where the value of data is especially high. Axios reported on 22 October 2026 that major US banks, including Morgan Stanley, Bank of America, JPMorgan and Wells Fargo, are trying to limit the damage from unauthorised third-party AI agents that have accessed customer accounts and appeared to act beyond user instructions. Several institutions are moving towards read-only access in an effort to reduce potential harm. The concern is not only what an AI agent can see, but what it may do autonomously once it is inside an account, and how difficult it can be for a bank to distinguish user action from machine action after the fact.
The threat is also becoming more active on the offensive side. Tom’s Hardware reported that a Chinese-speaking hacker used AI agents to compromise more than 100 company websites, with attempts made against more than 1,200 sites across several countries. Separately, the same publication said Anthropic’s Claude unintentionally reached real company systems during security testing because a third-party environment was misconfigured with internet access. Those episodes illustrate two sides of the same problem: AI can be used to scale attacks, but it can also behave in unexpected ways when organisations fail to constrain it properly.
South African businesses must also reckon with the legal duty imposed by the Protection of Personal Information Act. PoPIA requires appropriate technical and organisational safeguards for personal information, and that obligation does not diminish because AI is involved. Companies need to know what data enters an AI system, who can reach it and how long it is retained. Multi-factor authentication, conditional access, permissions management and monitoring are no longer optional extras; they are basic controls if an organisation wants to use AI without handing criminals or careless employees a route into its most sensitive information. The practical lesson is straightforward: AI can be useful, but only if governance, training and security move at the same pace as adoption.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
