As AI agents are given more access and autonomy, testing needs to go beyond whether they follow instructions. Katie Barnett looks at recent incidents involving OpenAI and other AI agents, and what they tell us about permissions, controls and testing what an agent is actually capable of doing.
Katie Barnett, Toro’s Director of Cyber Security, was quoted in SC Media UK following incidents involving AI models from OpenAI and Anthropic during security testing. The article looks beyond the “rogue AI” headlines to examine what the incidents mean for CISOs, including agent permissions, sandboxing, monitoring and the need for greater visibility over how AI agents are being used.
In this piece, Peter Connolly argues that AI is no longer just a tool but an active part of the workforce. As organisations adopt more autonomous, decision-making systems, the real challenge is not deployment but governance. Without clear accountability, identity controls and behavioural guardrails, digital agents can introduce risks similar to highly privileged insiders. His message is straightforward: treat AI like part of the workforce, with defined ownership, oversight and limits, or risk letting autonomy outpace control.
