Artificial intelligence (AI) agents are already active inside your business. But are you monitoring and managing them effectively?
This question was posed by Samantha Rule, CISO at Ninety One, during her presentation at the Cape Town leg of ITWeb Security Summit 2026, held yesterday at Century City Conference Centre.
“To be clear, I’m not just talking about chatbots that can answer FAQs. I’m talking about AI agents that have autonomy and that can act based on the information we give them.”
According to Rule, as AI agent deployment accelerates across enterprises, security and oversight are struggling to keep pace. She shared that the average enterprise has about 37 deployed agents, but only 24% have full visibility of agent-to-agent traffic. As such, it’s unsurprising that 88% of organisations reported a confirmed or suspected AI agent security incident in the last year.
Rule also shared that between 41% and 44% of organisations have no human-in-the-loop control over high-risk agent decisions, creating a growing security gap.
To mitigate this risk, Rule shared three key pieces of advice. Firstly, she encouraged organisations to treat agents like people. It won’t be long before a software developer has a team of agents working for them, which means agents will become an extension of the workforce and must be treated as such, she said.
“If a human used to do something and now an agent is doing it, the same identity, authority and accountability rules must apply.”
Her second piece of advice is to provide clear proof for every privileged action. “I am currently sitting with this problem, which is why I wanted to come up with a strategy to improve governance around the use of agents.”
This includes identity, authority, intent, chain of custody and accountability. “If you can’t answer who is acting (identity), on whose mandate (authority), to do what (intent), triggered by what (chain of custody) and reviewable by whom (accountability), you are putting your trust in something that cannot be verified,” she said.
Finally, Rule encouraged the audience to record all actions and install gates to safeguard against irreversible actions. If an agent can carry out actions that cannot be undone, like deleting data or transferring funds, it’s important to have policies in place that require human approval before the action can be executed.
Share
