49% found an AI agent with more access than it should have. Take that to Web Summit 2026

Web Summit opens in Lisbon on Monday 9 November and runs to Thursday. Many of the launches on stage next week will be AI agents: software that reads, decides and acts inside a company's systems. A demo shows what an agent can do. It cannot show what the agent should be allowed to do in your company, because only your company can decide that.

What happens after the demo

According to a survey of 300 US security leaders published by Frontegg on 27 October, 49% had found an agent running in production with access it should never have had. 78% had pulled permissions from an agent that was already live. Only 37% limit an agent's access to the task it was built for.

The survey covers companies with 500 or more staff, fielded by Cint in September. Frontegg sells tools for managing agent access, so the figures come from a vendor with a stake in the answer. They still describe a common sequence: the agent goes live with broad access, someone finds out what it can reach, and the access is cut back afterwards.

The decision belongs to the business

Each permission an agent holds stands for a decision someone made, or forgot to make: which records it may read, which actions it may take alone, and which wait for a person. When nobody writes those decisions down before launch, the agent's defaults make them instead, and the security team finds out later.

The people who know the right answers are the ones who do the work today. They know which refunds need a second look and which customer notes stay private. Their judgment is what the agent needs, and it has to be written down before the agent starts, in a form someone can check.

When nobody writes an agent's permissions down before launch, its defaults make those decisions instead.

The question to bring back

For any agent you see at the summit, from a vendor or from your own team, ask one question: which decisions will this agent take, and who wrote the rules it follows? If the answer is the vendor, or nobody yet, the work of deciding is still ahead of you.

AI strategy is one of the services Wyrework offers. We help leadership teams decide where agents should act and which decisions stay with people, before the first agent goes live.