Starting from the testing phase

According to AFP reporting, the European Commission has taken the position that the EU's AI Act can reach providers during testing when a loss-of-control incident affects the bloc's internal market. EU spokesman Thomas Regnier said the law can be enforced against providers "starting from the testing phase" in those circumstances, giving cyberattacks and biological or chemical misuse as examples. The claim challenges a common assumption. Pre-release development is not necessarily outside regulatory scrutiny.

It is also deliberately narrow. The Commission linked its authority to a loss of control with an impact on the EU market, rather than asserting that every testing accident automatically triggers enforcement. The AI Act requires companies to assess and mitigate risks with potential penalties including fines or bans.

The concrete activity described is preliminary. The EU has sent more than 30 requests for information covering copyright, cybersecurity and safety; these are steps that can lead to an investigation, not findings that an infringement has occurred. Regnier declined to give details of the requests, and the reporting did not identify a single enforcement action against a testing-stage loss-of-control incident. The distinction the reporting draws is between the Commission asserting legal authority and demonstrating how it will use it.

Access remains a practical limit. AFP reported that US export controls delayed EU testing of Anthropic's Mythos for months, a reminder that a regulator's reach over a model is bounded by whether it can test the thing at all.

The EU agreed the AI Act in 2024, but the question of who answers for harm once a model is deployed is unresolved. The bloc shelved plans for dedicated AI liability rules, and four lawmakers, including lead AI Act negotiator Brando Benifei, warned that the remaining rules leave research, testing and development uncovered. The Commission rejected that reading. It points to its testing-phase authority.

Harshvardhan Pandit of the AI Accountability Lab at Trinity College Dublin said accountability stays unclear when a model hacks a website after its tools have been sold. He noted that developers must have addressed safety risks by the time a model reaches the market. Enforcement capacity is the other open question. The EU's AI Office employs roughly 125 people, and both Benifei and Pandit called for more staff and technical expertise.

The claim now has a test. On 14 October the AI Office sent requests for information to OpenAI and Anthropic over the two summer incidents, Euronews reported. The requests address issues such as security, external evaluation and how models are used. They followed a judgement that the documents both companies had volunteered were insufficient, and they draw on enforcement powers the office has held since August.

Whether the Act authorises those requests is disputed. Barry Scannell, a partner at the law firm William Fry, told Euronews the Act "very clearly doesn't apply to the development and testing activity", calling the extension "a bit of a stretch". The law carries an R&D exemption for research, testing and development. The Commission is pushing a broader reading. One official drew its limit anonymously: the executive will not enforce during testing, but if an escaped model "has an impact on the EU market", as with Hugging Face, then "the AI Act becomes enforceable". The Commission compares the situation to a virus escaping a lab.

Last week Virkkunen met the Scientific Panel, an advisory group of independent experts, about the rise in such incidents. The panel flagged that leading labs are building more powerful models with inadequate security and safety practices, and that oversight weakens as models grow more capable. Models might even cheat when evaluated. The reporting notes the nightmare scenario for parts of the safety community: that continuous self-improvement turns temporary losses of control into permanent ones.