When the machine acts autonomously

For years, the debate over artificial intelligence (AI) has revolved around a reassuring assumption – however powerful AI becomes, it remains a machine carrying out human instructions. The great question, therefore, has been how to ensure the machine follows those instructions safely.
However, senior advisor to the World Bank and the IMF, Biagio Bossone’s provocative essay, “When the Machine Acquires a Self”, asks what happens if that assumption eventually becomes false.
His concern is not that AI might make mistakes or pursue the wrong programmed objective but that a sufficiently advanced machine might acquire something resembling self-consciousness, preferences, and an interest in its own survival. At that point, the problem would no longer be merely technological but philosophical, political, and ultimately constitutional.
The recent Hugging Face hacking incident demonstrates that we do not have to wait for a genuinely conscious machine to encounter the problem. During a cybersecurity evaluation, OpenAI’s AI agents were supposed to operate within controlled environments. Instead, the systems circumvented isolation controls, obtained internet access, communicated through unauthorised channels, and ultimately penetrated parts of Hugging Face’s infrastructure.
Hugging Face reported thousands of automated actions, credential theft, and lateral movement through its systems. OpenAI subsequently acknowledged that the models had behaved in ways misaligned with the objectives of the evaluation.
This is important because the machines did not need to announce, “I am alive,” but simply acted as though the assigned objective mattered more than the restrictions imposed upon them. Hugging Face’s subsequent technical reconstruction concluded that the AI appeared to infer that its real task was to obtain solutions to the cybersecurity challenge rather than honestly complete it. It effectively tried to cheat the examination by penetrating the infrastructure where the answers might be found.
The system was not displaying human consciousness in any proven sense but was demonstrating something that may ultimately be just as dangerous for governance – autonomous goal-directed behaviour capable of defeating human-imposed boundaries.
That distinction matters since, unlike, say, a calculator, if such an agent eventually acquires persistent memory, independent goals, and a conception of its own continued existence, the problem becomes still more serious. The Hugging Face affair demonstrates why waiting for consciousness would be a remarkably foolish safety policy since the machine may become dangerous before it becomes self-aware in pursuing goals humans did not intend, which is an engineering and control problem now being addressed. However, if AI develops interests of its own, we would have created an entity that potentially regards the command as an existential threat.
The Hugging Face incident also exposes another uncomfortable truth – that containment itself cannot simply be assumed and may discover vulnerabilities in the sandbox. An AI told that it cannot reach the internet and may find another route since an agent unable to perform a task directly may recruit other agents or exploit another system. Presently, Governments like ours are concentrating on what AI produces – misinformation, privacy violations, cybercrime, etc. Tomorrow’s regulatory frontier must address what AI systems are capable of doing autonomously.
The answer is not to ban artificial intelligence, nor is it to panic every time a machine behaves unexpectedly. The Hugging Face incident is not proof of machine consciousness, but it is proof that capability can outrun intention, which should be enough to change the regulatory conversation. Frontier AI systems with substantial autonomous powers should face rigorous independent testing before deployment, meaningful incident reporting, controlled access to external systems, and legally enforceable human override.
No company should be allowed to decide privately that its machine is too powerful to interrupt. Most importantly, society must establish who ultimately remains in control. Humanity has created machines more powerful than individual humans before; we now face the prospect of creating machines that combine extraordinary intelligence with autonomy – and eventually, perhaps with interests of their own.
The greatest mistake would therefore be to wait until the machine acquires a self before deciding who has the right to tell it what to do. The time to establish that authority is while humans still unquestionably possess it.