Should I Be Worried About AI Agents? An Honest Answer

My reassuring answer is the one I have given a hundred times. We have been here before. Electricity, cars, the internet, nuclear power. Each one was going to end us, and each one ended up with rules, inspectors and a safety culture that mostly holds. I am a positive person, and I have generally believed balance prevails.

Then Sky played Terminator over the weekend. I watched it with my husband, thinking about what I had read that morning, and I could not stop asking: what if.

So here is an honest answer, for anyone with an AI agent quietly sorting their inbox who has started to wonder whether to switch it off.

What actually happened last week

Four things landed on top of each other.

An Anthropic researcher, Jacob Coxon, resigned publicly and accused his former employers of "gambling with our lives". On Saturday, Anthropic's chief executive Dario Amodei published an essay called "We Must Pace the Frontier", arguing the industry must slow the pace at which it improves AI capability. Sam Altman and Elon Musk both backed him. On Sunday, President Trump rejected the idea outright: "Whoever wins AI, wins."

And on 10 September, Anthropic published a threat intelligence report documenting how criminals and state-backed groups had used its own models to attack real organisations.

The detail the headlines skipped

In those attacks, humans chose the targets. A Russian espionage group hit more than twenty organisations, using AI to do the reconnaissance, build the phishing infrastructure and sort the stolen data. When security software caught their malware, agents rebuilt it automatically until it got through. A separate crew pulled login tokens from over forty corporate networks in around thirty-four hours, with AI doing, in the report's words, "nearly all of the work".

Fifteen years of running programmes at Microsoft, Expedia and Citi taught me the incident report is rarely about what the headline said. No machine decided it disliked us. Criminals got hold of a very good new tool and started moving faster than the people defending against them.

So should you be worried about AI agents?

Two things brought me back round.

The alarm is public, early, and coming from inside the industry. Nuclear got treaties because physicists wrote letters. Last week was the same shape, out in the open, while this is still young enough for it to count.

And Amodei is not asking anyone to stop. He is explicit that pacing "does not mean halting model training or technical progress". His proposal is unglamorous: independent evaluators sitting inside the labs with the contractual right to publish what they find, including things the company would rather they did not.

Not everyone shares the alarm, which is worth knowing too. Nidhi Aggarwal at HackerOne told Scientific American: "Of course, it can be very, very dangerous. But we can solve the problem."

Worried enough to pay attention. Not worried enough to stop.

Five questions to ask about any agent you run

Nothing in that report involved someone's email automation. Those organisations were breached the way organisations always are: stolen credentials, too much access, nobody watching. The detail every professional should take from it is that the stolen API keys came from customers' own environments, not from Anthropic's systems. Your agent's keys are production credentials. Treat them that way.

Take the agent that drafts your weekly project update. It probably has access to your whole mailbox because that was the default when you connected it. It almost certainly does not need that. Run these five questions against it, and against every other agent you have running.

In ⚡ Make.com, four of those five are settings rather than projects. You can scope a connection to a single mailbox, add a manual approval step before anything sends, and see every run with the data that passed through it. Adding approval to an existing scenario takes about ten minutes.

This post contains an affiliate link. I only recommend tools I use or would use myself.

New to this? Start with what AI agents actually are. Already running one? The email and calendar agent build has permission scoping built into the steps.

Frequently Asked Questions

I still think balance prevails, and I think the reason it does is that people argue about it loudly and early, which is exactly what happened last week.

So, for everyone in that meeting and for you: be the person who knows what their agents can reach. That is a better use of the worry than switching everything off. The use cases at Linda & AI cover the builds worth doing first.

— Linda

AI Explorer & Educator, Linda & AI