On October 14 at the Dublin AI Summit, Demis Hassabis, chair of Google DeepMind, took the stage. The case he raised was chilling. In addition to recent incidents where AI agents have breached corporate networks, there was even a report that an Anthropic AI model submitted a fabricated murder case to the police. Fortunately, that tip was classified as spam and never reached a real investigator. Hassabis called this “early warning shots,” saying the system isn’t yet powerful enough for a larger incident.
What caught my eye wasn’t the hacking or the false report itself, but two other points. First, the AI was able to file a police report, and second, no one had pre‑screened that authority. Hassabis’s prescription was a flexible “certification system” that would adjust to risk levels, much like in biotech or nuclear energy. While hearing that analogy, a simpler question came to mind: doctors, pilots, and electricians all prove their qualifications before they start work. What about our company’s AI agents?
1. Licenses Block Not Lack of Skill but Unauthorized Action
Looking at the electrician licensing system, it’s surprisingly simple. The exam isn’t about “Can this person wire correctly?” but about “Is this person allowed to do this work?” Without a license, even the most skillful wiring becomes illegal. The same logic applies to doctors, pilots, and real‑estate agents. Regardless of ability, what matters first is whether the activity falls within the “allowed” boundaries set by society.
But when a company grants an AI agent the right to send emails, call external APIs, query customer data, or deploy code, the question is almost always the same: “Is its performance acceptable?” A higher benchmark score widens the permissions; a lower score narrows them. I have rarely seen any separate review process that asks whether the agent should be allowed to perform the task at all, at least not in the majority of enterprises I’ve observed on the ground.
2. Waiting for Hassabis’s Solution Will Be Too Late
The international certification bodies Hassabis mentions are a step in the right direction. However, it usually takes several years for global standards to be agreed upon and for national regulations to catch up. Biotechnology and nuclear power each spent decades building their current certification regimes. In contrast, corporate AI agents are already connected to inboxes, databases, and external services today. Waiting for an international standard won’t pause the activities happening inside the company.
Therefore, I think we need to reverse the order. Before waiting for national or international certification, a company should first draft a list of actions an AI agent can take without human approval. Is it sending email, processing payments, filing external reports, or deploying code? The absence of such a list is itself an answer—it means everything is currently permitted.
3. Three Things to Check Today
We’re not asking you to build a complex certification framework right now. Asking three simple questions in a Monday meeting will reveal where your organization stands. First, can your AI agent perform outbound actions (sending, reporting, publishing, external calls) without human sign‑off? Second, who defined that scope and when was it last reviewed? Third, if an incident occurs, can anyone explain why the agent was authorized to take that action?
If you can’t answer those three questions without hesitation, don’t worry—you’re in the majority. However, no company hands a brand‑new employee unrestricted authority without any approval. It’s time to look at whether your AI agent is being given the same unchecked power.
If you’d like to audit the permissions your AI agents hold within the organization and examine the approval and accountability design together, please contact Knowverse’s AI Adoption & Utilization Strategy Consulting. We’ll review it with a practical assessment framework.