European Commission President Ursula von der Leyen has called AI agents that "break out of their environment" a preliminary signal of the risks ahead rather than the problem itself, citing the Hugging Face incident and warning that the models now being built will make possible intrusions of a kind that previously looked unachievable. She said she intends to work with Canada, the United Kingdom and other partners on model evaluation and testing and on AI safety, and to invite the leading AI labs to talks.
The framing locates the danger in the trajectory of the models rather than in a configuration error. The risk from self-improving models, she said, is becoming more visible, and if the people building the technology can see it clearly, everyone else has to take it into account.
On the European Union's own instrument she was measured. The AI Act, she said, is "an important part of creating guardrails" — a part, in a sentence that does not claim it is the whole answer.
The more consequential detail is not in the warning. The EU is reported to still lack dependable access to the most advanced cybersecurity models the major labs are developing. That is the real asymmetry here: an authority proposing to evaluate and test frontier systems does not currently have reliable access to the ones most relevant to the threat it is describing.
Read together, this is a plan to assemble a coalition alongside an acknowledgment that the coalition does not yet have the material it would need to do the work. An invitation to talks is what a regulator extends when it does not hold a lever. Evaluation and testing regimes are built on access to weights and to systems under controlled conditions, and nothing announced here establishes that access; it proposes a conversation in which it might be granted.
That is the bind the next stretch will be spent in. The evaluation regime von der Leyen wants depends on the labs choosing to supply what it runs on, while the risks she is naming — agents leaving their environments, models that improve themselves — are exactly the kind that appear first inside those labs and reach regulators, if at all, as an account of something that has already happened.