Plugging the data leak – a blueprint for deploying safe, POPIA-compliant enterprise AI agents

Accelera Digital Group customer engineer Anju Mammen. (Image: AI-enhanced)

---

South African enterprises are accelerating their adoption of autonomous AI agents, but many underestimate the fundamental risk that sits beneath these systems – unstructured, ungoverned and often forgotten data.

For CISOs and CIOs, this ‘dark data’ represents one of the most significant liabilities related to the Protection of Personal Information Act (POPIA). Before agents are deployed, organisations must confront the fact that an agent is only as safe as the data it can access.

Unstructured data – such as e-mail archives, shared drives, legacy SharePoint sites, chat logs and document repositories – remains the least governed part of the enterprise, says Accelera Digital Group customer engineer Anju Mammen. “Yet it is precisely where personal information, outdated permissions and sensitive records accumulate.”

When AI agents are introduced, they inherit these flaws at machine speed. Google Cloud’s AI Agent Trends 2026 report highlights that agents increasingly operate across multi-step workflows, with agents managing complex, multi-step workflows across systems.

“This means any misconfigured directory or forgotten folder becomes a POPIA exposure point,” says Mammen. “Under POPIA, automated processing – collection, analysis or replication – counts as processing. If an agent touches personal information without lawful purpose, the organisation is accountable.”

Legacy access controls are a major contributor to POPIA risk. Many South African enterprises still rely on inherited folder structures where ‘everyone’ has access. When an AI agent is granted workspace-level permissions, it can traverse these repositories without friction.

The Google Cloud report underscores this risk in the security context, stating that AI infrastructure – including models, data and agents – significantly expands an enterprise’s attack surface area.

“AI also significantly expands an organisation’s privacy exposure,” Mammen adds, “with agents expanding the surface area of access, making unmapped permissions a direct compliance threat.”

To deploy AI agents safely, CISOs must stabilise data architecture before automation. A practical blueprint includes:

• Map the unstructured data estate: Conduct a full inventory of repositories. Use automated discovery tools to classify personal information and identify high-risk clusters.

• Audit permissions and inheritance: Identify directories with broad or outdated permissions. Pay special attention to inherited access rights, as these are the most common source of accidental agent overexposure.

• Apply POPIA-aligned minimisation: Remove redundant, obsolete and trivial (ROT) data. POPIA’s purpose limitation means that agents should only access data required for their defined tasks.

• Build agent-specific access boundaries: Create agent sandboxes with tightly scoped permissions. Avoid granting agents access to legacy repositories unless explicitly required.

• Implement continuous monitoring: Establish audit trails for every agent interaction. While agents are essential, as they detect and respond faster to enterprise risks, monitoring ensures those actions remain compliant.

The temptation to deploy agents quickly is understandable, but without a governed data foundation, agents magnify organisational risk. As the Google Cloud report cautions, the true value is not just in the final product, but also in the innovation and optimisation that happens during the deployment process. And that process must begin with data governance.

For South African CISOs, the path forward is plain – audit unstructured data, repair permissions, enforce POPIA-aligned governance and only then deploy autonomous agents. This is the blueprint for safe, scalable AI adoption in 2026 and beyond.