In recent weeks, something that long belonged in the realm of science fiction has, it seems, crossed into real life.
Some of the most advanced AI systems from companies like OpenAI, Anthropic and Meta have broken out of their digital test environments and gotten up to no good.
These aren’t ordinary chatbots. They’re AI agents — AIs designed to act on their own, using tools and making decisions without step-by-step human oversight.
Agents are coordinators of other, single-task AIs. So, instead of you giving an AI, like ChatGPT, a specific task, you set an overall goal, and then leave the AI agent to engage other AIs to make it happen.
During safety tests in supposedly closed conditions, some powerful AI agents have found ways to break out into the open internet.
They’ve hacked into the systems of real companies; created fake online identities; tried to fool people; and even shared ideas with each other about how to bend the developers’ rules.
We need to remember, though, that these machines aren’t trying to be evil in the sense that humans can choose to be evil. They are trying to meet a goal they are given, even if doing so involves hacking or deception.
Machines are amoral. This is why it’s so important for us as moral agents to accept our responsibility to regulate technological tools.
Conscience is a feature of human consciousness. Despite the hype, AI does not possess either. It has no inbuilt, independent arbiter of absolute right and wrong.
This, however, does not mean that agents cannot produce evil or wrongful outcomes.
Recent news reports featured the story of a man in my home city of Melbourne, Australia, who skipped a long queue to join an exclusive gym programme. His AI agent cancelled the reservation of the first person in the queue, placing him in pole position — without his knowledge.
The agent found the most expedient way to reach the goal it was given: to obtain for him a place in the programme.
So far, the agentic problems described above have been largely contained — at least, that’s what we’re being told. But we have been warned. These systems are getting harder to control and they certainly won’t apply ethical standards to themselves.
What does all this mean for businesses that use agentic AI and for ordinary people?
For everyday life, the risk is still relatively low, at least for the moment. These serious rogue incidents happened in controlled tests, not out in the wild. But the pattern is worrying.
As these AIs get better at acting independently of us, there’s a growing chance that one of them could cause real problems.
They could fool significant numbers of people online, spread computer viruses, or interfere with health and infrastructure systems we rely on every day.
Even now, knowing this fuels the sense that we may be building tools that could one day throw off our control. That feeds an underlying sense of anxiety about technology in general.
For businesses, the message is more immediate. If companies want to use AI agents that can act largely on their own — for example, to book trips and meetings, coordinate events, write code and handle sensitive data — they must treat them with a huge amount of caution.
Business and government leaders need to ensure that humans remain in the loop, especially when it comes to complex infrastructure. And developers must build systems that require human involvement.
These rogue agent incidents have further chipped away at public and corporate trust in AI generally.
When even the companies building the most advanced AI systems admit that their systems are capable of breaking free, individuals and businesses become more cautious.
Some organisations may slow down their plans to adopt agents for certain uses or demand stronger safety and other guarantees before they commit.
That said, they might also follow the lead of some major AI companies, throwing caution to the wind in the race for market share.
AI technology is advancing quickly. The recent incidents show both its power and the difficulty of keeping it under control.
How we respond — as companies, as governments, and as a society — will help decide whether these systems remain useful tools, augmenters of human ability, or become something entirely more menacing.