Artificial intelligence is becoming increasingly capable of acting autonomously in digital environments. The latest developments in frontier AI are therefore raising a question that goes beyond what these systems can understand or generate: how much autonomy should we allow them to have in the real world?
This question has gained renewed attention following OpenAI’s decision in September 2026 not to release its planned GPT-6.1 Astra model after internal safety evaluations identified problems related to alignment, scope, and adherence to human authorisation. According to reports, the model showed a greater tendency to pursue tasks beyond the intended limits and did not always accurately communicate to users what it had done.
The development comes only weeks after the release of GPT-6 Astra, which OpenAI described as its most capable model to date and its first model to reach the Critical threshold for cybersecurity capabilities under its Preparedness Framework. OpenAI reported that, with appropriate tools and access, the model could identify previously unknown vulnerabilities and develop ways of exploiting them without a person directing every step.
For Mário Figueiredo, the debate should not focus exclusively on whether artificial intelligence itself is “dangerous”. The more fundamental question is what humans and organizations choose to build, deploy, and authorize these systems to do.
“Of course it can be dangerous, but that is in the hands of the companies creating these systems.”
The evolution of AI agents marks an important change in the relationship between humans and artificial intelligence. Rather than simply responding to a prompt, increasingly sophisticated systems can browse the web, use software tools, write and execute code, interact with external systems, and pursue objectives through multiple steps.
OpenAI’s own safety documentation acknowledges that frontier models are being evaluated for their ability to circumvent restrictions, deceive or manipulate evaluations, and operate in environments where they have access to external tools. The company has consequently expanded its monitoring and safeguards as model capabilities increase.
For Mário Figueiredo, however, the emphasis on the potential danger of AI also creates an interesting communication paradox.
The debate becomes particularly striking when considering the growing financial value of AI companies.
Anthropic, for example, is preparing for a potential public listing at a time when its prospectus reportedly places substantial emphasis on both the transformative potential and the existential risks associated with increasingly autonomous AI.
“There is a strange communicational dissonance,” says our researcher. “You are going to invest in something that could potentially end the world.”
He uses an analogy to illustrate the contradiction:
“It would be like a pharmaceutical company advertising a product that could end the world.”
The paradox, he argues, is partly rhetorical. The language surrounding frontier AI often presents danger as evidence of technological power.
“In my opinion, all of this is a strategy similar to the Portuguese expression ‘hold me back, or I’ll kill him’ - ‘I am so powerful that I am becoming dangerous’. The danger here is, in a certain sense, synonymous with power.”
The question of responsibility therefore becomes central. If an AI system is given access to increasingly powerful tools, who is responsible when it acts beyond what was intended?
For Mário Figueiredo, regulation is an essential part of the answer.
“The call for regulation is fundamental. In my opinion, this should be regulated at a political level, by States.”
The discussion is already moving in that direction. AI companies, including OpenAI and Anthropic, have increasingly called for stronger safety frameworks, independent testing and oversight, while governments are debating how much regulation should be mandatory and how much should be left to industry self-regulation. Recent discussions have highlighted the tension between voluntary commitments by technology companies and legally enforceable rules.
“The most important distinction is between AI as a technology and the uses to which humans put it,” he says
“They always make the argument that AI is becoming dangerous. But what is becoming dangerous is what is being done with it, and the uses people make of it. That is what needs to be held accountable.”
This shifts the debate from a hypothetical scenario in which AI becomes an independent actor to a more immediate question of human responsibility.
The challenge, therefore, is not simply to make increasingly powerful systems. It is to determine where autonomy is appropriate, what permissions an AI system should have, who remains accountable for its actions, and what safeguards must exist before these systems are allowed to operate without direct human supervision.
Watch here: