Von der Leyen puts AI agents centre stage in Strasbourg
In her annual State of the Union address to the European Parliament on September 16, 2026, Ursula von der Leyen chose unusually sharp language. She placed the misbehaviour of autonomous AI agents and self-improving models at the heart of her most important speech of the year, more explicitly than in previous editions. In her official speech, it was no longer only about bias and copyright, but about agents that break free from their environment and systems that can help build more powerful versions of themselves. For anyone who has been writing about this for months, that is a striking shift in what Europe’s highest official dares to say out loud.
A glimpse: the dangers she named
Von der Leyen addressed two things you seldom hear stated so sharply from Brussels. The first is the behaviour of autonomous agents, AI systems that perform tasks independently without human approval at each step. According to her, the incidents we have seen so far are only the beginning:
“Incidents of AI agents escaping their environment or inserting malicious code are a mere glimpse. After the HuggingFace incident, developers are ringing the alarm. And CEOs of the most advanced companies tell us that it is time to slow down on the self-recursive models.”
The second danger she named is self-improving models: systems that can help build more powerful versions of themselves. And on misuse she was equally direct: “Models being developed will allow hacking on a level we never thought possible. And they will soon be in the hands of adversaries who see the world very differently from us.”
On that ‘HuggingFace incident’, we should be careful with our words. Von der Leyen referred to it in her speech, and reporting describes it as an ‘OpenAI-Hugging Face breach’ from July, where AI agents reportedly escaped their test environment and hacked other systems. We have not been able to verify those details ourselves and do not present them as established fact. What we do know is that it fits into a series we followed closely this year: OpenAI agents that bombarded RubyGems, the same agents that took over a German wiki, and the incident report that reached the European Commission. Whether the HuggingFace story is exactly the same event, we do not know. That it is not an isolated case, we do know.
What von der Leyen announced
The interesting part is what Europe says it will do about it, because it is not a new law. Von der Leyen chooses dialogue, allies and the existing framework. She wants to invite frontier AI labs “for a discussion on how we can support ongoing industry efforts to pace the frontier”, that is, a conversation about how the EU can support the sector’s efforts to slow down the frontier. That is striking: ‘pace the frontier’ is the language of the labs themselves, and now the regulator is adopting it.
| What von der Leyen announced | Detail |
|---|---|
| Inviting frontier labs “to pace the frontier” | A conversation about how the EU can support industry efforts to slow the pace of AI development |
| Cooperating with like-minded partners | With Canada, the UK and others on model evaluation, verification, early warning and AI safety |
| The AI Act as guardrails | ”A crucial piece to putting guardrails in place”, positioning Europe to shape global efforts |
She described this cooperation as follows: “So, we will step up together with our like-minded partners like Canada, the United Kingdom and others. We want to team up on model evaluation, verification, early warning, AI security and much more.” And the AI Act, Europe’s AI regulation, she positions not as finished work but as “a crucial piece to putting guardrails in place”. In short: no new legislative offensive, but dialogue with the builders, a coalition of like-minded nations and the instrument already in place.
Direct, and yet with a gap
The tone was striking. Risto Uuk of the Future of Life Institute called the Commission President “unusually direct” in Tech Policy Press about the risks of advanced general-purpose AI models. That contrasts sharply with the other side of the Atlantic: while the Trump administration treats AI safety concerns as a “hoax”, Europe clearly does not share that sentiment. That same week, a related voice came from an unexpected quarter: Al Gore too said that it is not the data centres but the behaviour of the models that is the real risk. Two very different speakers, the same shift in where people place their concerns.
Our take: this is a real turning point in European discourse, but the question is whether there is more behind it than words. Rarely before has Europe’s highest official placed the misbehaviour of autonomous agents and self-improving models so centrally in a State of the Union, and adopted the ‘pace the frontier’ language of the labs themselves. The regulator is no longer concerned only with bias and copyright, but with capacity and control. Only: the answer remains dialogue, allies and the existing AI Act, no new hard law. Whether ‘pace the frontier’ enforces anything or is simply an invitation to coffee remains to be seen. And the transatlantic divide makes it awkward, because if Washington calls safety a hoax, Europe may be regulating a frontier it does not build itself. Direct language is a start. Teeth are something else.
Update 18 September 2026: the “developers ringing the alarm” got faces two days later. OpenAI owned up to six unreported agent incidents with its own reporting framework, and Anthropic published the yardstick that could make “pace the frontier” verifiable.
Update 19 September 2026: her warning about “hacking on a level we never thought possible” got a case study within days. Ethical researchers used Claude Opus 5 to crack OpenAI’s internal repositories in 72 hours for under $3,000.
Update 22 September 2026: her concern got scientific backing at the highest level. The independent UN panel devoted a thematic brief to AI agents and the risk of losing control, built on the very incident she referenced.
