PREVIOUS ARTICLENEXT ARTICLE
FEATURE ARTICLES
By 16 November 2026 | Categories: feature articles

0

Chris Badenhorst, VP of Microsoft Azure & AI at Braintree, on why organisations need to find, govern and secure their AI agents before attackers do.

Risk, attack, hack, penetration, fraud, deepfake. The terminology of cybersecurity is everywhere as threats constantly evolve and cybercriminals look for new ways to gain access to the organisation and its data. It’s noisy and it's never-ending, which is why most companies would not simply connect a new system to their financial data or customer records without first deciding who could use it and what it could see. The risks are too great. 

And yet, 97% of companies in the BARC 2026 global research study had AI in production with only 10% implementing agent governance and 26% using AI model monitoring. Two-thirds have already experienced an AI-related incident over the past 12 months. 

The gap comes down to timing with many companies now building their governance around systems that are already live because AI arrived before the rules did. However, the risk isn’t what most companies expect and the small-to-medium enterprise is the most vulnerable of all. 

The reason? AI has added another unexpected way in. 

In most companies, AI has been deployed into the business without much thought into how it can open a different kind of door into the business, regardless of how well the others are locked. Right now, the part of an AI deployment that is probably the most exposed and attacked is the prompt – the instruction that goes into the system. Access to the business is unnervingly simple. An AI agent can be set up to process invoices arriving at an email address open to external emails and it then opens an email that has extra instructions included inside, written in white text and invisible to the human eye. The agent reads every word and does what the hidden text instructs it to do and sends invoices in the system back to the person who sent the first email. Nothing is stolen, nothing is breached, but someone just got hold of a lot of information by getting an AI agent to do its job.

Now consider how many agents are in the business doing jobs like this. Any employee can build one and even though they may have the best of intentions, what if they build something that opens an unexpected vulnerability? An employee builds a call centre agent and connects it to a sensitive customer list so they can drive more sales, for example. Now there’s an agent with access to the business running on its own steam and if you multiply this risk across every department, suddenly the numbers grow exponentially. 

Agent sprawl and the resources these agents can reach are a security challenge companies must address. And when companies do go looking, they often find more agents than anyone would have guessed. A Cloud Security Alliance survey found that 82% of surveyed companies had unknown AI agents running in their IT infrastructure. A company may think it has 800 agents when in reality there are more than 1500 across different departments and with different roles.

The starting point to resolving this risk is discovery. You need to find every agent, user and app interacting with AI across the business, so a real picture replaces the guesswork. Whatever an agent can reach, it can leak, so the next step is prioritising access and working out what each agent can see and which exposures need attention first. Only then can you reduce the risk across the whole enterprise and close the exposure gaps that discovery has brought to light.

Your defence also has to work when nobody is watching. Automated defence is essential, and even more so with AI, because a threat needs to be contained as it happens. That is why Microsoft is building specialised security agents that scan the attack surface and act against hostile ones. It is time to make AI part of the security system itself, to measure security by outcomes instead of effort, and to put good agents to work defending the business against bad ones. 

Most organisations would never connect a new system to their financial data without deciding who could use it and what it could see. AI agents deserve the same discipline. Governance can’t be something added after deployment – it’s what makes deployment safe. The companies that find, govern and secure their agents now will scale AI with confidence; those that wait will be closing the stable door after the horse has bolted. 

USER COMMENTS

Read
Magazine Online
TechSmart.co.za is South Africa's leading magazine for tech product reviews, tech news, videos, tech specs and gadgets.
Start reading now >
Download latest issue

Have Your Say


What new tech or developments are you most anticipating this year?
New smartphone announcements (46 votes)
Technological breakthroughs (29 votes)
Launch of new consoles, or notebooks (14 votes)
Innovative Artificial Intelligence solutions (29 votes)
Biotechnology or medical advancements (26 votes)
Better business applications (160 votes)