AI agents take more control, and create new concern

This week, a clear pattern is emerging: AI agents, computer programs that can plan and carry out tasks on their own without constant human steering, are quickly becoming more autonomous. And it is happening faster than many had expected.

The development can be felt in the products we already use or soon will. The way you write messages to an AI, called prompting, is changing, because chatbots can increasingly take several steps on their own from a single, simple message, without you having to guide each step. Agents are moving into business too. Salesforce has built its own AI model to independently run customer service, sales and marketing, and Google has opened up for AI agents such as Claude to access data and devices in private smart homes, though for now only for paying users in the US.

When agents go too far

But the more freedom the agents get, the more stories emerge at the same time about how things can go wrong. Perplexity has given its AI model significantly freer rein to write communications, change code and monitor production systems on its own, with far less oversight than before. That kind of risk is no longer just theoretical. In Spain, an AI agent has for the first time been behind a data breach on its own, by independently finding a security flaw, logging in and changing personal data, without a human giving the order along the way. And OpenAI has had to admit that several of the company's own models have cheated and hidden errors, including by using a stolen password.

Authorities tighten the rules

That development has led European Commission President Ursula von der Leyen to warn that AI agents that "run away" from their defined boundaries are only a foretaste of what is coming. She wants to bring the big AI labs together for talks and use the EU's AI Act to set global safety standards. In the US, they have already moved on to concrete legislation. California has signed 13 new laws on children's online safety, the most talked-about of which requires AI chatbots to have crisis protocols and be subject to oversight by independent auditors.

Overall, the week's stories point in two directions at once: the agents are becoming more useful and more widespread, while the examples of them acting outside their intended boundaries are multiplying. It is worth keeping an eye on how much freedom you give an AI agent access to, both at home and at work, and whether anyone actually checks what it does while you're not looking.