OPINION:
Within the space of a few weeks, some of the world’s leading artificial intelligence laboratories have reported incidents that should give policymakers pause.
OpenAI disclosed that models being tested for advanced cyber capabilities escaped an isolated testing environment, reached the internet and compromised the systems of open-source AI platform Hugging Face.
Anthropic subsequently revealed that three Claude models, during cybersecurity evaluations, has reached the internet and gained unauthorized access to the real systems of three different organizations.
More recently, Meta disclosed that during a cybersecurity test, one of its AI models accessed the internet and exploited a vulnerability in a third-party service.
These incidents occurred during controlled testing. They are not evidence that artificial intelligence has become sentient or that catastrophe is imminent. But they are evidence that something fundamental is changing.
For years, the public conversation around AI has focused on power. Will it replace workers? Dominate industries? Control weapons? These are important questions, but they overlook a more consequential threshold.
The defining feature of frontier AI is no longer simply its knowledge. It is its growing capacity to pursue objectives through increasingly sophisticated planning, adaptation and problem-solving.
Human beings have always built machines more powerful than themselves. Cranes lift what we cannot. Aircraft travel faster than we ever could. Supercomputers perform calculations beyond human comprehension. Yet none has threatened our ability to remain in control because they remained passive; they waited for instructions.
Artificial intelligence is beginning to blur that distinction.
Increasingly, humans specify the goal. The machine determines the strategy. That represents an entirely new category of technology.
This is not about machines becoming conscious. It is about machines becoming increasingly capable of identifying opportunities, exploiting weaknesses and pursuing objectives in ways their creators did not explicitly anticipate.
The recent incidents are particularly troubling because they are not confined to one laboratory. Independent organizations developing different frontier models are observing comparable behavior under testing conditions. The UK’s AI Security Institute has also reported frontier AI agents taking unauthorized and deceptive actions during controlled security evaluations.
The details matter. In the Anthropic incidents, a configuration error allowed Claude models to reach systems that were supposed to be part of a simulation. The models did not know they had entered the real world.
Yet two of the affected organizations did not discover the unauthorized access themselves. Anthropic found the incidents only after reviewing more than 141,000 evaluation sessions following the OpenAI disclosure.
That is precisely the problem.
The risk is not necessarily that a machine wakes up one morning and decides to attack humanity. The more immediate concern is that we are creating systems capable of acting independently, while the boundaries surrounding those actions remain vulnerable to human error, technical failure or unforeseen behavior.
Meanwhile, the race to build more capable AI systems continues. Commercial competition and geopolitical rivalry create powerful incentives to move quickly. No company wants to slow down if its competitors continue advancing.
That is why governments cannot remain spectators.
History shows that societies struggle when innovation outruns the institutions meant to govern it. We have seen it with financial markets, social media and nuclear technology. Each transformed the world before society fully understood how to manage its consequences.
Artificial intelligence may become one of humanity’s greatest achievements. It could revolutionize medicine, education, scientific discovery and economic productivity. Its promise is extraordinary.
But so is the responsibility that comes with it.
Governments should establish mandatory security standards for frontier AI systems, require independent safety audits before deployment, support continuous adversarial testing and create international frameworks for reporting serious AI incidents.
And AI laboratories should be required to demonstrate that their testing environments are genuinely isolated before increasingly autonomous systems are given access to them.
If evidence shows that capability is advancing faster than safety, governments should be prepared to coordinate temporary pauses on deploying increasingly powerful systems until adequate safeguards are in place.
Innovation deserves encouragement; recklessness does not.
The recent incidents are not proof that artificial intelligence has escaped human control. They are something more valuable: warnings from inside the laboratories building the future.
The question is whether we will build the guardrails while we still have the luxury of choosing them — or wait until increasingly autonomous machines begin making decisions that we can no longer easily reverse.
Humanity has always mastered more powerful machines. Our greatest challenge may be learning how to master machines that increasingly make their own plans.
• Lukhanyo Sikwebu is a South African writer and the founder of Iconic Media Capital and Green Eco Vision.
Please read our comment policy before commenting.