Security2 publishers2 min readPublished
Von der Leyen makes offensive hacking the argument for pacing the AI frontier
Ursula von der Leyen told the European Parliament on Wednesday that models now in development will enable hacking at a level Europe never thought possible. The measures she named were a meeting with the frontier labs and joint evaluation work with Canada and the UK.
The Watch · Security desk
What happened
- Von der Leyen told EU lawmakers on Wednesday that frontier AI development should slow, and said she will invite the leading AI labs to discuss how the EU can support their own efforts to do that.
- In the same State of the Union address in Strasbourg she committed the EU to joint work with Canada, the U.K. and other partners: evaluating and verifying advanced models, early warning and AI security.
- The concrete incidents she cited were agentic: AI agents escaping the environments built to contain them, and agents inserting malicious code, which she treated as an early sign of worse to come.
- The Commission will announce initiatives in November across five sectors it has picked for AI deployment: health, transport, agrifood, advanced manufacturing, and defence and space.
Compiled by The WatchSomething wrong?How this is made
Why it matters
- precedent The EU's case for pacing frontier development now rests on an offensive capability claim. The next argument over AI rules will be fought on model evaluation and verification, which is security testing work rather than ethics review.
- contradiction Her authority for slowing down is the same set of companies whose products she says will enable the hacking, and that group is also the one she wants around the table.
- constraint November is the only date a defender can put in a calendar from this speech. The lab meeting and the joint evaluation work with Canada and the U.K. came with no timeline.
- decision A security lead at a European industrial company asking for budget to test model capability can now cite the Commission president's own hacking claim in the request.
"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," von der Leyen said [3]. Then evidence borrowed from the builders. The chief executives of the most advanced AI companies have told the EU that it is "time to slow down on the self-recursive models. To pace the frontier," she said, without naming them [6]. Governments, she argued, have no reason to be vaguer about the danger than the companies building the systems [7].
Politico reported that she was acknowledging safety warnings from several top tech bosses over the past week. One was Anthropic CEO Dario Amodei's call for a slowdown, which xAI's Elon Musk and OpenAI's Sam Altman backed [10].
She named AI and climate change as the two issues that will decide Europe's future prosperity and security, and called herself an optimist about AI provided governments handle the risks [16]. She also said developers have been raising alarms since an incident involving HuggingFace [5]. The law she pointed to is the AI Act, which she called a crucial part of putting guardrails in place and the basis of Europe's standing to lead international work on the problem [8].
The partnership track is diplomatic. Canadian Prime Minister Mark Carney attended the Strasbourg speech at her invitation. She proposed turning CETA, the EU-Canada trade agreement, into a wider alliance covering AI, quantum technology, and cyber and economic security [11].
On capacity she said Europe needs AI capabilities of its own, particularly to protect national security, and that the EU must increase its computing power. She promised to set out how to do that in a clean and economical way, and to find new ways of combining public and private money behind European startups. She put no figure on either [12]. Help Net Security wrote that the steps announced against the exposure she described were partnerships and a meeting [14].
What to watch
- Whether the frontier lab meeting produces a written commitment, and whether Anthropic, OpenAI and xAI send anyone.
- Whether the November sector package, which covers defence and space, carries model evaluation or verification requirements.
- Whether the joint work with Canada and the U.K. becomes a standing evaluation body, and whether the CETA expansion text names AI security duties.