AI agents have no common sense: why data-level control matters

Organisations are moving AI agents out of pilots and into production faster than they are putting controls around them. In this Cyber Pulse interview, Sam Harrison, Channel Manager EMEA at Kiteworks, explains why agents create a new kind of data risk, why visibility alone is not enough, and why the answer is to protect the data itself.

Key highlights

  • AI agents carry out tasks with whatever access they are given, without the judgement a person would apply.
  • In IBM's 2025 research, 97% of organisations that suffered an AI-related breach had no proper AI access controls in place.
  • Visibility is the foundation, but control has to sit on the data: who is asking, and how sensitive the data is.
  • Sam's advice: treat AI agents as digital employees, with the same oversight and limits as human staff.
  • Kiteworks applies role- and attribute-based policies to every file request, including requests from AI assistants.

The race to production has opened security gaps

The push to adopt AI for productivity has moved faster than governance. IBM's Cost of a Data Breach 2025 found 13% of organisations reported breaches of AI models or applications. Of those, 97% had no proper AI access controls, and 63% of breached organisations had no AI governance policy or were still writing one.

Sam puts the core problem simply: agents are given a task and a level of access, and they execute.

"AI agents act without common sense. You would like to hope that a human being would employ [it] if they're being asked to send sensitive financial data to a Hotmail email address. But if AI is able to do it, then AI will do it."
— Sam Harrison, Channel Manager EMEA, Kiteworks

When AI becomes part of the attack surface

The risk is not only an agent making a mistake. Attackers can also turn the AI against its owner. In June 2025, researchers at Aim Security disclosed EchoLeak, a critical flaw in Microsoft 365 Copilot. A crafted email could trick the assistant into leaking internal data, with no click from the user. Microsoft patched it, and there was no evidence it was exploited in the wild, but it showed how an AI with broad data access becomes a path out of the organisation.

Visibility is the foundation, not the finish line

Many organisations believe that seeing their data is enough. Sam agrees visibility matters, but says AI changes the pace at which control has to work.

The approach he describes ties policy to the data, not only to the user. Kiteworks' policy engine looks at the user's role and at the attributes of the data: what it is and how sensitive it is. New data, whether created internally or brought in from outside, is tagged, and policy follows it.

"It doesn't matter if the AI agents are moving fast … because we're protecting the data itself."
— Sam Harrison, Channel Manager EMEA, Kiteworks

That protection applies to people as well as agents. The aim is to stop both from leaking data by accident.

Treat AI agents like employees

When AI becomes part of the attack surface, Sam's advice is to go back to fundamentals. Security exists to protect the organisation: its intellectual property, its reputation and its data. Attackers want the data, and that has not changed.

"What we have to do is treat our digital employees, our agents, the same way we treat our human employees, in that we need to make sure that we have visibility over what they're doing and control over the data that they're accessing and what they're doing with it."
— Sam Harrison, Channel Manager EMEA, Kiteworks

How Kiteworks puts control on the data

Kiteworks' Private Data Network governs how sensitive data moves between people, systems and AI. For AI specifically, its Secure MCP Server lets AI assistants work with company files while governance rules still apply:

  • Role-based and attribute-based access controls, enforced by the Data Policy Engine using real-time context.
  • Policies that follow data classification labels.
  • Audit logs of every interaction, fed into SIEM for compliance reporting and investigation.

Amiviz and Kiteworks in the region

Amiviz and Kiteworks have partnered since March 2025 to bring Kiteworks to organisations across the Middle East and Africa, with a focus on finance, healthcare, energy and government.

What this means for MEA organisations

As agents move into production, the first question is no longer only who can access sensitive data, but what can. A practical starting point: list the AI agents and assistants already connected to company data, give each one only the access its task needs, make sure sensitive data is classified, and log what agents do with it. Controls that sit on the data keep working however fast the agents move.

Put guardrails around your AI agents

Want to see how Kiteworks governs what AI agents can access and share? Book a demo through our Kiteworks page, email info@amiviz.com or contact us.

‍