Sponsored

Samresh Ramjith, Executive Head: Enterprise Security, Vodacom Business

Artificial intelligence is transforming cybersecurity with a speed and intensity that few organisations are prepared to manage. It is strengthening attackers, expanding the digital attack surface, and challenging established security controls while giving defenders powerful new methods to identify threats, process complex telemetry, and respond at machine speed. Organisations must rethink how they govern technology, protect information, manage identities and allocate security investment to adopt AI safely while managing real and perceived risks.

According to telemetry and survey data presented in INTERPOL’s 2026 assessment, AI was involved in 55% of surveyed cybercrime cases in 2025. Reported losses increased from US$192 million in 2024 to US$484 million in 2025. Partner telemetry attributed 92% of recorded African ransomware detections and almost 40% of phishing detections to South Africa.

Attackers have embraced AI faster than many organisations can respond. The attraction is simple: AI changes the economics of cybercrime by reducing the cost and expertise required for sophisticated attacks. An attacker no longer needs a large technical team to research a target, imitate an executive’s communication style, translate fraudulent messages or generate malicious-code variants. Generative AI can perform these activities rapidly, at scale and at a fraction of the traditional cost, collapsing the window between vulnerability detection and exploitation.

Synthetic speech, deepfake video and realistic branding can create the appearance of authority and may be difficult to distinguish from genuine communications without advanced detection and forensic capabilities. Fraudulent messages and other forms of social engineering remain significant initial-access mechanisms, reinforcing the importance of protecting human interactions as well as technical infrastructure.

AI systems also introduce vulnerabilities and dependencies, including exposure to prompt injection, data poisoning and modelinversion attacks in poorly designed or protected systems. A distinctive risk is that AI systems can fail without an obvious technical error. A generative model may, for example, produce a confident but false output, commonly described as a hallucination. Several organisations have faced criticism after publishing material that proved inaccurate, fabricated or misleading because users relied on AI-generated output without adequate validation. Human-in-the-loop or human-on-the-loop oversight is therefore necessary for consequential uses.

The AI supply chain introduces risk through dependence on external models, datasets, libraries, plugins, cloud platforms and specialised providers. Agentic AI intensifies these by moving from content generation towards autonomous execution. A compromised chatbot may produce an inaccurate answer that becomes an operational action. With excessive privileges, poor supervision or sensitive-system access, a manipulated instruction could affect multiple processes at machine speed. Traditional application controls remain necessary but may be insufficient without controls governing agent identity, authority, supervision and containment.

A common response is to block public AI platforms or prohibit their use. Blanket restrictions are unlikely to succeed and may drive employees towards unapproved services on personal devices or subscriptions. A better strategy is to create secure pathways for responsible adoption, supported by governance that enables innovation while setting clear boundaries for data use, system access, human oversight and accountability.

The first requirement is to gain visibility of existing AI usage. Organisations need an authoritative inventory of their AI models, agents, datasets, identities, APIs, and external dependencies. Security teams cannot govern AI assets that they cannot identify, and boards cannot evaluate exposures that remain hidden within business units or individual workflows.

The second requirement is strong identity and access management. AI agents should be treated as non-human identities with controlled permissions, protected credentials and traceable activity. Access must follow least privilege, giving an agent only the information and capabilities required for its approved function. High-impact activities involving finance, regulated decisions, customer rights, safety or critical infrastructure should require meaningful human authorisation at appropriate stages. The reviewer must have the expertise to challenge AI-generated output and the authority to escalate or halt the process.

The third requirement is secure engineering across the AI lifecycle. Data should be classified, protected and traceable, with models, code and supporting components rigorously tested before deployment. AI outputs can be probabilistic, context-sensitive and difficult to predict consistently. Performance can change as data, context and usage evolve, even when the model is not continuously retrained. Continuous monitoring is therefore needed to detect drift, degradation and behaviour outside approved boundaries.

AI’s major defensive opportunity lies in its ability to correlate large volumes of security telemetry, identify anomalous patterns and draw on external intelligence to contextualise and respond to those patterns. In a region affected by shortages of specialist skills and uneven investigative capacity, these AI-enabled investigative and cyber-defence capabilities could materially improve the productivity of security teams. INTERPOL specifically recommends AI-assisted detection, anomaly analysis, deepfake identification, proactive threat hunting, and stronger AI literacy, while retaining human judgement for consequential decisions.

The strategic winners will govern AI before scaling it, embed security into architecture, maintain meaningful human control and use AI to augment scarce cyber expertise. In the AI era, cybersecurity should not brake innovation but evolve to operate as the control system that allows it to accelerate without compromising trust, resilience or accountability.

References

Boston Consulting Group. (2026, August 27). Cybersecurity budgets are growing fast: AI threats are growing faster.

Deloitte. (2025). The cognitive leap: How to reimagine work with AI agents. Deloitte Consulting LLP.

Deloitte. (2026, April 24). Agentic AI is scaling faster than guardrails. Deloitte Insights.

International Criminal Police Organization. (2026). African cyberthreat assessment report 2026. INTERPOL.

International Organization for Standardization. (2023). Information technology: Artificial intelligence: Management system (ISO/IEC Standard No. 42001:2023).

MITRE. (n.d.). Adversarial threat landscape for artificial-intelligence systems: MITRE ATLAS. https://atlas.mitre.org/.

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0)

(NIST AI 100-1). U.S. Department of Commerce. https://www.nist.gov/itl/ai-risk-management-framework

OWASP Foundation. (2026). OWASP Top 10 for large language model applications 2026. OWASP Generative AI Security Project.

Vodacom Business, & Omdia. (2024). Cybersecurity for growth: Protecting South African organisations in an evolving digital environment. Vodacom Business.

vodacombusiness.co.za