Five practical checks for AI agents with the power to act

As AI agents move from making recommendations to changing systems, security leaders need to govern what they can do, not just what they can access. Cybersecurity executive Phillimon Zongo sets out five controls, including clear limits on autonomy and rules that trigger a fresh review when an agent’s powers grow.

Watch Desk analysis

What happened

In a Forbes Councils article published on 28 October, Zongo argues that one-off approval is too weak for AI systems whose permissions, reach or decision-making authority can expand. He recommends ongoing governance built around the consequences of a system getting something wrong.

Key findings

• Classify what the AI is allowed to decide
  Record whether it may recommend, approve, change, trigger or execute an action, and the potential business impact at each level.
• Start with credible ways it could fail
  Consider inaccurate decisions, manipulated inputs, corrupted data and service outages, then connect those scenarios to likely business consequences.
• Scale controls with autonomy and impact
  A recommendation may need less oversight than an agent able to affect production systems or multiple business units. Zongo argues that greater authority and a wider blast radius should bring stronger controls.
• Set triggers for renewed review
  Reassess a use case if the system gains autonomy or privileged access, its potential impact grows, its dependencies change, or its actions become harder to reverse.
• Make oversight forums decide something
  Put unresolved risks, proposed controls, residual exposure and the person authorised to accept it in front of executives, rather than burying decisions in status updates.

Why it matters

These are practical questions for organisations moving agents from pilots into consequential work. For example, an AI security tool that recommends disabling an account presents a different decision from one authorised to do it automatically. The same system may need a fresh review if it later gains access to production servers.

Zongo’s advice is that controls should follow the potential harm, not merely the name of the technology. That gives security teams a way to distinguish a low-risk experiment from an agent whose mistake could disrupt critical operations.

Our read

The strongest point is that “human oversight” is not a magic spell. A person needs time, expertise and authority to intervene, and the system needs clear review triggers when its remit changes. Start by documenting what an agent can actually do and what happens if it gets it wrong. The rest of the governance conversation should follow from that, rather than from a checklist looking for a home.

Zongo presents these as recommendations, not evidence that every organisation follows them or that they guarantee safe outcomes.

What to watch

• Whether organisations define review triggers before expanding an agent’s permissions.
• Whether human reviewers can stop or reverse actions, rather than merely flag concerns.
• Whether executive governance meetings focus on unresolved decisions and named risk owners.

Discussion spark: When an AI agent gains new permissions, should that automatically trigger formal reapproval, or should teams be able to expand its role under existing controls?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.