Securing the AI super-user: Why cyber security must lead SA’s agentic future

Daniel Acton, Chief Technology Officer at Accelera Digital Group (ADG).

---

As South African enterprises accelerate their adoption of AI agents, a critical shift is under way, combining opportunity with unprecedented risk. These agents are no longer simple assistants; they can act by processing transactions, accessing sensitive client files, initiating workflows and interacting with core business systems. In effect, they are becoming “super‑users”.

“The moment an AI agent can take action inside your organisation, its access to systems and data are the same as any other user’s, and the agent should be treated accordingly. This will ensure you maintain control over its actions,” says Daniel Acton, CTO at Accelera Digital Group (ADG).

This is not a hypothetical concern. The Cybersecurity Forecast 2026 report highlights that threat actors are already exploiting AI systems, with prompt‑injection attacks expected to rise sharply as adversaries target enterprise AI deployments for data exfiltration and sabotage. The report notes that “targeted attacks on enterprise AI systems” will increase as attackers move from proof‑of‑concept exploits to large‑scale operations.

At the same time, the Cybersecurity Forecast report shows that organisations are rapidly adopting agentic systems across workflows – from finance to customer service to security operations.

However, with this adoption comes a new class of risk, that of shadow agents. These are autonomous agents deployed by employees without oversight, creating invisible pipelines for sensitive data. The report warns that this will escalate into a critical challenge by 2026.

“South African businesses cannot afford a repeat of the shadow IT era. This time the stakes are much higher. An unsanctioned AI agent is not just a compliance risk; it is a potential new attack vector,” Acton notes.

Traditional identity and access management were designed for humans. AI agents challenge that model, operating continuously, interacting with multiple systems and executing multi‑step workflows at machine speed.

The Cybersecurity Forecast anticipates the rise of “agentic identity management” – a new discipline where AI agents are treated as distinct digital actors requiring granular and dynamic access controls.

“If an AI agent can approve a transaction, read a client file or trigger a workflow, it must be governed with the same – or greater – rigour as a human super‑user. Anything less is an open invitation to cyber criminals,” says Acton.

South Africa consistently ranks among the world’s most targeted regions for cyber attacks. Combined with stringent local compliance requirements (such as the Protection of Personal Information Act (POPIA), Financial Sector Conduct Authority (FSCA) regulations and King IV), enterprises face the dual challenge of rapid AI adoption and rising cyber risk.

The report highlights that adversaries are increasingly using AI to scale attacks, automate reconnaissance and exploit human weaknesses through AI‑enabled social engineering.

“Attackers are already using AI to move faster. If defenders don’t build AI securely from day one, they’re fighting a losing battle,” says Acton.

Acton argues that the only sustainable path forward is a security‑by‑design approach, which entails embedding governance, identity controls and monitoring into the data architecture before deploying agents.

He states that the path forward is clear – secure the data foundation, govern the agents and build AI systems that enhance the organisation without exposing it. Those who prioritise security today will be the ones who unlock the full promise of AI tomorrow; safely, confidently and at scale.

“AI will transform South African enterprises, but only if we secure the foundation. Governance is not a barrier to innovation; It is the enabler that makes innovation safe. The winners in 2026 won’t be the fastest adopters but rather the ones who adopt responsibly,” Acton concludes.