AI Security in South Africa

AI security in South Africa means securing and governing your AI and LLM systems against both global AI standards — the EU AI Act, the NIST AI Risk Management Framework, ISO/IEC 42001 and the OWASP Top 10 for LLM Applications — and local expectations, including POPIA. CyberSigma threat-models your AI stack, tests guardrails and data flows, reviews governance, and hands you a prioritised, board-ready report. We are CERT-In empanelled and PCI QSA (CEMEA) authorised.

Which AI security standards and rules apply in South Africa?

AI security sits at the intersection of fast-moving global AI standards and South Africa's own posture. South Africa is developing national AI policy, while POPIA already governs the personal data that AI systems process. An assessment is only useful if it is scoped to what actually applies to your use cases. The ones we most often map to:

• EU AI Act — the risk-tiered AI regulation that applies extraterritorially to anyone placing AI systems on the EU market or whose AI output is used in the EU; high-risk and general-purpose AI carry specific obligations.
• NIST AI Risk Management Framework (AI RMF 1.0) — the leading voluntary framework for governing, mapping, measuring and managing AI risk.
• ISO/IEC 42001:2023 — the certifiable AI Management System standard, with ISO/IEC 23894 for AI risk management.
• OWASP Top 10 for LLM Applications — the de-facto checklist for securing LLM and generative-AI features.
• Your data-protection law — because AI systems process personal data, POPIA applies to training data, prompts and outputs.

What a CyberSigma South Africa AI security review actually covers

We assess how your AI systems behave under attack and how they handle data, not just your model card. In a typical engagement we:

• Threat-model the AI/LLM stack against the OWASP Top 10 for LLM Applications — prompt injection, insecure output handling, sensitive-information disclosure, excessive agency and the rest.
• Test guardrails for real: jailbreak and prompt-injection resistance, input/output filtering, and what the model does with untrusted content from tools, RAG sources and users.
• Review data governance for AI: what personal or sensitive data feeds training, fine-tuning and prompts, the lawful basis for it, and whether customer data leaks across tenants or into a foundation model.
• Secure the model pipeline: access to model registries, secrets and MLOps tooling, supply-chain risk in third-party and foundation models, and protection against model and data exfiltration.
• Assess governance and oversight against ISO/IEC 42001 and the NIST AI RMF: accountability, human oversight, documentation, and evaluation of high-risk use cases.
• Check monitoring and incident response for AI-specific failure modes — abuse, drift, harmful output — not just classic infrastructure alerts.

Representative engagement: a Johannesburg enterprise

A useful way to picture the work: a Johannesburg enterprise adopting AI assistants needed assurance around data protection and misuse. We red-teamed the systems against the OWASP LLM Top 10, reviewed governance, and mapped data flows to POPIA, then handed them a prioritised fix list. This example is representative of how we structure South Africa AI security reviews; named client references are available under NDA on request.

How long does an AI security review take, and what does it cost?

Most AI security reviews run a few weeks, depending on how many models, applications and data flows are in scope. Cost follows that scope rather than a fixed list price, so we run a short, free discovery call, agree the scope in writing, and give you a fixed quote before any work starts. If you are working to a launch, customer or regulatory deadline, tell us the date and we will tell you honestly whether it is achievable.

Does the EU AI Act apply to companies in South Africa?

It can. The EU AI Act applies extraterritorially: if you place an AI system on the EU market, or the output of your AI is used in the EU, its obligations can reach you even when you operate from South Africa. We assess which risk tier your use cases fall into and what that means for you.

What is the difference between AI security and a normal penetration test?

A classic pen test looks at infrastructure and application flaws. AI security adds the model layer — prompt injection, jailbreaks, training-data and RAG poisoning, sensitive-data disclosure, excessive agency in agents, and model/data exfiltration — plus the governance and data-protection questions AI raises. We do both and join them up.

Do you test LLM and generative-AI features specifically?

Yes. We threat-model and test against the OWASP Top 10 for LLM Applications — including prompt injection, insecure output handling, supply-chain and plugin risks, and overreliance — for chatbots, copilots, RAG systems and AI agents.

How does POPIA affect our AI systems?

AI systems usually process personal data, so POPIA applies to training data, prompts and outputs — lawful basis, transparency, data-subject rights and cross-border transfers all come into play. We assess your AI data flows against it as part of the review.

How often should we review AI security?

At least annually, and again whenever you ship a major model change, adopt a new foundation model, add an AI agent with new tool access, or change what data feeds the system.