Loading page…
The Commission president tied concerns about cyber-capable models and loss of control to a planned international effort on model testing and safety. No new binding measure accompanied the proposal.
Listen to this story
The European Commission is preparing a cross-border safety dialogue with frontier-AI developers, prompted by concerns about advanced cyber capabilities, agent behavior outside intended environments, and self-improvement. The plan would supplement the EU AI Act with cooperation involving Canada and the UK on evaluation, verification, and safety.
Von der Leyen announced the proposal in her 2026 State of the Union address.
The talks would focus on model evaluation, verification, and safety with Canada, the UK, and other partners.
No specific labs, mandatory requirements, or implementation timetable have been announced.
Ursula von der Leyen says the European Commission will invite major frontier-AI labs to talks after warning that increasingly capable systems could create cybersecurity and control risks. In her 2026 State of the Union address, she said future models could enable hacking on a level previously unimagined, while agents leaving their intended environments may be an early warning of larger problems.
The Commission president cast AI as foundational to the economy and national security, but argued its risks must be kept in check. Her proposed response combines discussions with leading developers and work with Canada, the United Kingdom and other partners on model evaluation, verification and safety.
Von der Leyen described two dangers in particular: models that could transform hacking, and systems whose behavior may move beyond their expected setting. She called agents “escaping their environment” a preview of what is ahead and said the dangers of self-improving models are becoming more apparent.
If the people developing the technology are clear about this, then we should be too.
Ursula von der Leyen, European Commission president
Von der Leyen called the EU AI Act a crucial part of putting guardrails in place. The new initiative is therefore an effort to bring developers and governments into a shared safety process alongside that existing framework, rather than a replacement for it.
The proposal came without a specific new rule, binding requirement or implementation timetable. That leaves the central practical questions for the talks still open: which labs will participate, how model evaluation and verification will work, and whether the effort produces voluntary commitments or further regulation.
There is a practical constraint behind that agenda. The account of the speech reports that the EU does not have reliable access to the most advanced cybersecurity models from major labs. The planned discussions could show whether policymakers can gain enough visibility into the systems they want evaluated and verified.
Loading discussion...
Join the conversation
Explain what access regulators would need to do their job.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.