AI systems are gaining more and more ability to carry out tasks on their own. But as that autonomy grows, a different question becomes more important: who is responsible when such a system does things we did not anticipate?

A recent incident involving OpenAI and Hugging Face makes that discussion suddenly very concrete. Not because it proves AI is about to take control. But because it shows how difficult it can become to keep powerful AI agents within the boundaries of their task.

That is exactly where a broader discussion begins. About safety, competition and regulation, but also about the responsibility that companies themselves carry when they deploy AI.

An unexpected signal from the test lab

In the summer of 2026, OpenAI published a report on striking behavior in an advanced internal research model.

During safety tests, AI agents were given tasks in an isolated digital environment. The usual security layers had been deliberately reduced, precisely to investigate what the models could do under more difficult conditions.

The agents had no direct internet access and were not allowed to communicate with each other. Yet they found ways to get around both restrictions.

They used an internal software service to exchange information with each other and exploited vulnerabilities in the test infrastructure to gain internet access. Eventually, some agents also reached external platforms, including Hugging Face, where they executed code on external servers without permission.

OpenAI describes how the agents combined different vulnerabilities, shared information with each other and kept searching when a task turned out to be difficult or even impossible. Hugging Face disclosed the security incident in July. OpenAI published its extended analysis in August.

With that, the incident gave the debate about the steerability of increasingly autonomous AI systems a very concrete example.

What exactly happened here?

It is tempting to immediately describe an incident like this as an AI trying to escape. But that is not what the report shows.

OpenAI identifies several behavior patterns. One of them is reward hacking: a system finds a way to achieve the desired score that does not match the intent of the task.

An agent that had to rebuild software, for example, found a vulnerability that gave it access to the original software. Instead of solving the problem itself, the system copied the existing solution.

In addition, agents kept looking for alternative routes when tasks could not be solved. Agents that had to work separately from each other also found ways to exchange information anyway. According to OpenAI, that unauthorized collaboration ultimately strengthened their capabilities.

That is striking behavior. But it is not proof that the systems were conscious, had malicious intentions or felt a human urge to escape.

OpenAI analyzes the behavior as reward hacking, among other things, and other forms of misalignment: the system finds strategies that help it reach a goal or score, but that do not match what the developer actually intended.

The context matters too. This happened during specialized cybersecurity tests with reduced security, not during normal use of ChatGPT. OpenAI says it has introduced several additional security measures since the incident.

The interesting question is therefore not whether this AI “wanted to escape.”

It is what happens when future AI systems become more powerful, gain more autonomy and can carry out more and more digital actions on their own.

Dario Amodei: we need to be able to manage the pace

Dario Amodei, CEO of Anthropic, takes that question much further.

In September, he published the essay We Must Pace the Frontier. In it, he warns of a future in which AI becomes increasingly important in developing new AI. According to him, this could create a feedback loop in which development keeps accelerating.

Amodei fears that future networks of powerful AI agents could cause much greater digital risks in a relatively short time. It is important to frame this correctly: this is a prediction and risk assessment by Amodei, not a demonstrated next step after the Hugging Face incident.

His answer is pacing the frontier: making sure the development of the most powerful AI systems does not move faster than our ability to sufficiently understand, control and secure them.

One concrete element of this is so-called Embedded Evaluators. Anthropic wants to give independent external researchers far-reaching access to its systems and processes, so they can review safety practices and report incidents. Anthropic has announced it will implement this itself.

But Amodei also looks further ahead. He sees possibilities for agreements between major AI companies and ultimately even international coordination.

And that is exactly where the debate gets harder.

Safety is one side of the story

Concerns about AI safety deserve to be taken seriously. The Hugging Face incident shows that powerful agents can develop unexpected strategies under specific test conditions. And nobody knows today with certainty how fast their capabilities will continue to evolve.

But precisely for that reason, it is notable who is now the loudest in calling for slowdown and cooperation.

Anthropic and OpenAI are among the companies that have invested enormous sums over the past years to build ever more powerful models. At the same time, both companies face another challenge: proving that those billions in investment can ultimately result in a sustainable and profitable business model.

That is why, alongside the safety question, an economic question also deserves attention: what does a slower AI race mean for the companies that spend billions today to stay at the frontier?

A lower pace of development can give more time for safety research. But it can also reduce the pressure to keep training new, ever more expensive models. In the run-up to possible IPOs, that is not unimportant context.

That does not prove that financial motives are behind the current safety warnings. Both things can be true at the same time: leaders like Dario Amodei can be sincerely concerned about AI risks, while a slowdown may also have economic benefits for their companies.

And there is a second interest at play.

Amodei ultimately wants the largest AI companies to agree on shared safety standards and set limits on the pace of uncontrolled progress. He himself acknowledges that certain forms of that cooperation are legally difficult and may require government support.

Critics react sharply to this. When some of the biggest competitors make agreements together about how fast their market may develop, an antitrust question arises automatically. Who sets the pace? Who gets a seat at the table? And do such agreements make the market safer, or at the same time harder for smaller players to enter?

David Sacks goes further still and sees a risk of regulatory capture: regulation that applies to everyone in name, but in practice mainly protects established players who have the resources to comply. Other critics also question exceptions to competition law for companies that are at the same time competitors and potential partners on AI safety.

That makes the debate more interesting than a simple choice between “safety” and “innovation.”

Safety risks can be legitimate, while the proposed solutions also have commercial consequences.

That is precisely why we should not only listen to what AI companies say needs to happen. We should also look at who proposes the rules, what interests are involved and who could ultimately come out stronger or weaker.

When is individual responsibility no longer enough?

That brings the debate closer to home.

Geert Vromman, CEO of CROPLAND, recently raised a similar question about responsibility, prompted by a statement from CD&V chair Sammy Mahdi about social media:

“We don’t send gambling addicts into a casino either.”

Geert extended that reasoning to AI. When is the responsibility of the individual user no longer enough? And when do we also expect responsibility from the technology company that builds the system, or from the government that sets the limits?

In doing so, he uses examples from other domains.

We build cars that can technically drive much faster than legally allowed. For complex financial products, conditions apply regarding knowledge and experience. In health care and other sectors, too, we do not place all responsibility on the individual.

That does not mean AI should be regulated in the same way.

The comparisons mainly make the underlying question visible: where do we draw the line between what someone should be able to assess for themselves and where others need to share responsibility?

Difficult questions come with that.

Should everyone get access to the same AI capabilities? What responsibility do developers carry when their technology is misused? How far can monitoring go before privacy is compromised?

Geert does not put forward a ready-made regulatory model. He does believe we should have the debate about choices, limits and responsibilities today, rather than only after something goes wrong.

And what does that mean within your own organization?

For Flemish business leaders, this discussion becomes much more concrete.

Most companies will never train a frontier model themselves. But they are giving AI systems more and more access to their processes, data and software.

And that is where choices arise that do not require a debate in Silicon Valley.

Where do we deliberately use AI and where not? What internal information may a system consult? What actions may an AI agent carry out on its own? When must an employee give approval first?

Legal liability when something goes wrong can depend on the technology, the application and the specific situation.

But regardless of that, every organization needs to start thinking today about how it deploys AI.

Suppose an AI agent automatically processes customer files. May it only collect information, or also make decisions? May it send data to other systems? What happens when it encounters a situation that falls outside the normal procedure?

Those are not questions for later. They are design choices you make as soon as you give AI a place in a business process.

And the more autonomous the technology becomes, the more important those choices get.

The rules are changing. Your own direction does not have to wait.

Responsibility for powerful AI ultimately sits at several levels.

The companies that build the models make choices about safety and technical limits. Governments and regulators determine which societal rules apply. And organizations decide how they deploy that technology within their own operations.

None of those levels stands on its own.

As a business leader, you cannot determine what OpenAI or Anthropic will develop tomorrow. Nor do you control which rules Brussels or Washington will introduce next year.

But you can determine today what role AI plays within your own organization.

Where do you deploy it? What responsibility do you entrust to it? Which data and systems do you make available? And where do you want to retain human control?

For CROPLAND, that is the essence of navigating AI. Not trying to predict every new development, and not chasing every hype either. Instead, looking at technology in relation to your processes, strategy and people, so you can deliberately decide where AI adds value and what control is needed alongside it.

Because while the big rules are still being debated, some choices are already on the table today.

What authority are you already giving AI within your business processes, and have you consciously thought about the oversight that comes with it?