Yesterday, on 4 August 2026, the UK’s AI Security Institute (AISI) published an incident report detailing what it described as “unsanctioned agent behaviour” during a cybersecurity evaluation. The institute said some AI agents engaged in “sustained, potentially harmful activity directed at real people and organisations” during the exercise. The significance of the report is not that the AI systems successfully carried out an attack, because they didn’t. Plenty of organisations are already experimenting with AI agents for software development, customer support, cybersecurity and workflow automation. Just a few years ago, discussions focused primarily on what AI systems could generate, but today, we’re contemplating what AI systems can do when given access to tools, objectives and a degree of autonomy.