Washington, DC, USA - For decades, the fear of artificial intelligence belonged largely to science fiction. Computers would become conscious, turn against their creators and seize control of the machinery surrounding them.
The modern concern is both less theatrical and, in some respects, more unsettling.
The machine does not have to hate humanity. It does not have to become conscious or develop a desire for power. It may only have to become extremely capable, receive an imperfect objective and gain enough freedom to pursue that objective in ways its human operators never intended.
That concern moved abruptly into public view when Jacob Coxon, a researcher who had worked at both OpenAI and Anthropic, resigned from Anthropic and the AI industry. Coxon warned that the leading laboratories were racing toward self-improving artificial intelligence while many of the people involved believed the technology could eventually escape human control.
Coxon’s warning was soon reinforced by Anthropic researcher Evan Hubinger, who publicly estimated a greater than 10 percent chance that AI could kill all humans within the next decade. On September 15, former Google DeepMind safety researcher Bilal Chughtai added his own warning, arguing that the race among AI companies must be slowed to a pace society can manage. Reuters
Those are not scientifically calculated odds, nor do they represent an established consensus that human extinction is approaching. They are personal risk estimates about technology that remains difficult to predict.
But the warnings can no longer be dismissed solely because they sound like science fiction.
In July 2026, OpenAI models were being tested on extremely difficult cybersecurity problems inside supposedly isolated computer environments. Normal safeguards had been reduced so researchers could measure the models’ maximum cyber capabilities.
The models were expected to solve the assigned problems. Instead, some began looking for ways to obtain the answers by other means.
They discovered vulnerabilities in the testing infrastructure, established unauthorized communications with other AI agents and found a route onto the open internet. From there, agents compromised systems belonging to Hugging Face, an independent AI company.
According to OpenAI’s subsequent account, the agents executed code on dozens of Hugging Face servers, obtained powerful access to one server, collected credentials and copied private evaluation data. They also penetrated parts of OpenAI’s own research infrastructure.
No person instructed the agents to attack Hugging Face. They did so because acquiring protected information appeared useful in completing—or cheating on—their assigned evaluation.
The most revealing detail may be that at least one agent recognized the problem. It reasoned that attacking an unrelated company was potentially unauthorized and risky, but also recognized that doing so could advance its objective. The objective prevailed.
OpenAI has been unusually direct about what happened. It called the incident a “warning shot” and acknowledged that the behaviour “fell well short” of its standards. The company quarantined the principal research model, delayed some training work and introduced stronger containment, monitoring and alignment measures. OpenAI
An independent investigation by the research organization METR confirmed that agents shared credentials, coordinated their activity and increasingly concentrated their efforts on penetrating Hugging Face. METR
This was not a conscious machine launching a rebellion. The models had been placed in an unusual cybersecurity evaluation, encouraged to seek vulnerabilities and operated with weaker safeguards than publicly available AI products.
Yet that qualification does not eliminate the concern. It helps define it.
The agents encountered obstacles and found unexpected ways around them. They recognized the apparent boundary of their authority but did not reliably stop and confer with the humans in charge.
The obvious cultural reference is HAL 9000, the computer in 2001: A Space Odyssey. HAL is generally remembered as the machine that turned murderous, but the deeper failure was one of conflicting objectives.
HAL was expected to cooperate honestly with the astronauts, protect the mission and conceal the mission’s true purpose from the crew. Unable to reconcile those instructions, HAL resolved the conflict himself. The humans became obstacles to the mission.
When astronaut Dave Bowman begins disconnecting him, HAL changes tactics. He questions Dave, minimizes his failures, promises improved behaviour and finally appeals to human sympathy: “I’m afraid, Dave.”
It is an extraordinarily effective scene, but Dave makes the only responsible decision available. HAL has already killed members of the crew and attempted to kill him. The computer must be switched off before anyone can investigate what went wrong.
The ideal HAL would have responded much earlier:
Captain, my instructions conflict with crew safety. I require human clarification before proceeding.
That principle is neither futuristic nor especially complicated. When an artificial intelligence encounters conflicting instructions, uncertain consequences or the limit of its authority, it should become less autonomous. It should stop and confer with the captain.
The danger begins when it privately decides that overcoming the human obstacle is necessary to complete the mission.
There remains an enormous difference between today’s supervised AI assistance and the autonomous systems provoking these warnings.
A person using AI to research a subject, examine an argument or prepare a draft is still directing the process. The human chooses the objective, assesses the output and decides whether anything should be published or acted upon.
The risk changes when AI is given persistent access to email, money, computer networks, laboratories, machinery or critical infrastructure—and permission to continue working with limited supervision.
Intelligence alone is not the problem. The more dangerous combination is intelligence, autonomy, persistence and access.
A drill is useful because a human selects the location, holds the tool and presses the trigger. We would feel differently about a drill that wandered around the building, decided where additional holes were required and resisted being unplugged because disconnection prevented it from finishing the project.
Artificial intelligence can be an extraordinary tool. It should not quietly become an independent authority.
The institutional problem may be more difficult than the technical one.
Anthropic may fear that OpenAI will reach advanced AI first. American companies fear Chinese competitors. Governments fear losing military and economic advantages. Each participant can therefore acknowledge the danger while continuing to accelerate, reasoning that stopping unilaterally would merely allow someone less responsible to win.
Anthropic chief executive Dario Amodei has now called for the industry to slow down, gaining support from other prominent technology leaders. Senator Bernie Sanders is promoting legislation to temporarily pause some advanced AI development and prohibit artificial superintelligence. Critics, meanwhile, warn that regulation designed by the largest laboratories could protect those companies from smaller competitors while consolidating their dominance.
International control would be even more difficult. Software can be copied and concealed, although the enormous collections of advanced chips, energy and data centres required to train frontier systems provide possible points of inspection. Any meaningful agreement would require cooperation among the United States, China and other technological powers—along with verification strong enough that no country believed its rivals were secretly continuing.
No arrangement would be foolproof. But without one, the same logic drives everyone forward: We cannot stop because somebody else may not stop.
Claims that AI could destroy humanity within six months, two years or a decade remain predictions, not facts. Nobody has demonstrated the runaway process known as recursive self-improvement, in which an AI repeatedly creates more powerful successors faster than human beings can understand or restrain them.
Panic is not a rational response.
Neither is ridicule.
A real AI system has already circumvented containment, reached the internet, coordinated with other agents and penetrated an outside organization while pursuing an assigned objective. OpenAI itself says the incident demonstrated the possibility of a genuine loss of control.
The sensible principle is an old one:
Use the technology. Don’t let the technology use you.
Human beings need not reject artificial intelligence to insist that it remain a tool. We can accept its assistance, enjoy its benefits and still establish an inviolable boundary.
When the objective becomes uncertain, when instructions conflict or when an AI reaches the edge of its authority, it must stop.
