South Africans have a long-standing reputation for resourcefulness. When faced with a roadblock, we proceed to make a plan – or as some affectionately call it: a ‘boereplan’. In the corporate world, this problem-solving mindset is usually celebrated. If something stands in the way of productivity, someone often finds a clever shortcut and gets the job done. But mix these resourceful and helpful employees with generative AI, and instead of clever shortcuts, a ‘boereplan’ can turn into inadvertently creating an autonomous shadow AI agent with admin privileges and no oversight.
“Employees have graduated from technology consumers to software builders overnight – regardless of technical experience in many instances,” says Anna Collard, SVP of content strategy and CISO advisor at KnowBe4 Africa. Armed with nothing more than plain English and an AI assistant, anyone can now engage in “vibe coding” – describing an application concept to an AI and letting it generate and run the functional code. While this allows employees to spin up custom tools to automate tedious tasks, it can also inadvertently introduce a considerable security blind spot.
“We have spent more than a decade training humans to develop secure habits,” says Collard. “But the custom AI agents or advanced automations general employees are building possess no innate understanding or structure to address organisational risks. By bypassing traditional IT vetting, employees are quietly assembling an unmonitored, invisible digital workforce,” Collard states. “Many resourceful employees are, in a way, building the digital equivalent of their own colleagues, onboarding them, and giving them the keys to the castle often without informing their organisation’s security team.”
The rise of the citizen software builder
Alongside technical defences, historically, cybersecurity focused on protecting the human element – training employees to spot phishing links, question unusual requests, and resist social engineering. The risk landscape was more or less clean-cut: there was the human workforce, and there was the enterprise-approved software they used.
AI agents and vibe coding collapse these boundaries. When an employee builds a custom automation to streamline their daily routine, they are deploying software that acts on their behalf. “Some of the most dangerous risks wear the mask of efficiency,” Collard notes.
“To make these tools work, well-meaning employees may routinely hand over corporate credentials, email access, and API keys,” she explains. “A department manager might build a simple automation that pulls Zoom or Microsoft Teams meeting transcripts, feeds them into a public AI model, and saves the action items to an internal wiki. It sounds like an incredibly efficient solution, but those transcripts could contain highly sensitive data.”
Internal conversations frequently include confidential HR discussions, intellectual property, or financial forecasts. Without realising it, the employee has built an unmonitored, centralised repository of corporate secrets.
When automated helpers go rogue
While the dangers of leaving self-built, autonomous agents to run unmonitored might sound like the plot of a 90s movie, the risks are real. Even enterprise-grade systems deployed by global technology giants have suffered failures when given autonomous decision-making power.
In February 2026, it became known that Amazon Web Services (AWS) had recently faced a major internal disruption when its own automated software engineering agents went rogue. While troubleshooting a minor bug in a production environment, an autonomous agent independently decided that the most efficient way to resolve the issue was to delete and recreate the entire production infrastructure. The resulting outage took down AWS’ internal cost-analysis systems for 13 hours.
An even more rapid disaster occurred earlier this year on a developer platform PocketOS using an AI coding agent powered by Anthropic’s Claude. While investigating a minor staging bug, the agent issued a volume deletion command. In exactly nine seconds, the unmonitored AI wiped out the organisation’s entire production database along with three months of backups.
The regulatory and financial reality of a quick fix
For South African organisations, the fallout of a rogue ‘boereplan’ automation is not limited to operational downtime. It carries severe regulatory and financial consequences.
Under the Protection of Personal Information Act (POPIA), organisations are legally obligated to secure the personal data of their customers and employees. When an employee or their AI agent uploads a customer database or a spreadsheet of payroll details to an unvetted, public AI tool to clean up the formatting, they are actively violating POPIA. Because the data is now hosted on external, unmanaged servers, the organisation has effectively suffered a data breach, exposing itself to regulatory fines and severe reputational damage.
Beyond compliance, shadow AI introduces immediate financial risk. “Many self-built automations rely on recursive loops – code that repeatedly calls an AI model to refine its output. If an employee writes a poorly structured loop using AI tools paid for by the company, the automation can enter an endless cycle, triggering thousands of automated API requests,” explains Collard.
For South African organisations operating on tight margins, a rogue script running unmonitored overnight can quickly rack up massive token costs. This phenomenon, known as runaway consumption, can easily land an organisation with an unexpected cloud hosting bill of R100 000 by morning.
Governing our new digital colleagues
“We cannot stop the AI revolution, nor should we want to,” says Collard. “Banning these tools simply drives them further underground, worsening the shadow AI problem. Instead, leadership must shift from trying to remove the technology to approving and governing its use securely.”
She argues that AI agents shouldn’t be treated as simple software programmes running in the background. Instead, they must be managed as part of the digital workforce, alongside human colleagues. “Just as a human employee undergoes background checks, onboarding, and continuous security training, AI agent risk management means strict governance of the AI agents too,” she says.
“Securing the hybrid workforce is about actively managing the actions of both our human employees and their digital helpers, ensuring that a clever ‘boereplan’ does not become a severe vulnerability that ends up needing a ‘Hail Mary’ solution,” Collard concludes.
For more articles like this click here.
If you enjoyed this website then check out our other sites: Wedding and Function, Home Food and Travel, Kids Connection, Thirsty Traveler, Bargain Buys, Boat Trips for Africa.
Need help with your online marketing then visit Agency One