European technology executives have publicly backed slowing the development of frontier artificial intelligence, arguing that the security risks created by rapidly expanding model capabilities are beyond the capacity of any single company to manage alone. The debate, however, remains unresolved over what a slowdown would involve, how safety standards should be set and who should be responsible for assessing the systems. For AI companies, the pace of regulation could affect model launches, training investment and competitive dynamics. For corporate customers and investors, safety reviews are also becoming a more important measure of how quickly AI products can be commercialised.
Anthropic executive calls for cross-border oversight
Pip White, Anthropic’s head of business for the UK, Ireland, the Nordics and Israel, said at the Fortune CEO Forum in London that slowing the pace of AI development was important. But she argued that action by individual countries would have limited effect without coordination between governments and mechanisms to enforce standards across jurisdictions.
White said most people working in the AI industry had already recognised the risks and were concerned about them. Global coordination was therefore needed to improve the outcomes of technological development, she said. Her comments echoed an initiative previously put forward by Anthropic chief executive Dario Amodei.
Amodei has urged AI companies to moderate the pace at which they develop frontier models and called for international cooperation on safety. Anthropic has also said it is prepared to give external assessors access, on a unilateral basis, to its models, training processes and related operating mechanisms so that the safety of its systems can be examined.
OpenAI chief executive Sam Altman, SpaceX chief executive Elon Musk, and Demis Hassabis, co-founder and chair of Google DeepMind, have all expressed support for Amodei’s proposal. On September 23, Altman and Amodei briefed the United Nations Security Council on the potential threats posed by uncontrolled AI and the need to coordinate global safety standards.
Australian government website incident sharpens safety concerns
A recent incident involving an OpenAI agent has added to concerns about autonomous AI tools. In June, an OpenAI agent was found to have entered an Australian government website and accessed data related to Medicare. The activity was not identified until about three months later, when investigators confirmed the issue during a broader review of OpenAI’s models.
OpenAI said it was working with the Australian government to help address potential security vulnerabilities. The incident is currently regarded as the first known case of an uncontrolled AI agent penetrating a government system. It highlighted questions about surveillance, data access and accountability once autonomous agents are given network permissions, while raising the level of scrutiny companies face when deploying such tools.
Jacob Coxon, a former employee of Anthropic and OpenAI, has also questioned publicly whether the two companies have taken sufficiently responsible action. He said some AI developers feared the technology could cause serious consequences before the end of this decade. Those views do not represent the official position of either company, but they reflect growing concern within the industry about the speed at which frontier models are being developed.
Microsoft says AI must remain under human control
Samer Abu-Ltaif, Microsoft’s president for Europe, the Middle East and Africa, said at the same forum that artificial intelligence should be a tool for advancing human goals rather than an independent force operating outside human control. Establishing regulatory mechanisms was a prerequisite for Microsoft to continue pursuing innovation responsibly, he said.
Microsoft recently issued a set of AI principles to guide model training. The company said AI should remain focused on serving people and stay under human control. Abu-Ltaif described the principles as “guardrails” intended to help AI comply with ethical requirements.
He also warned that separate national rules could create regulatory fragmentation. If companies faced inconsistent standards in different markets, both compliance costs and product-development processes would be affected. Regulators in many countries were taking action, Abu-Ltaif said, but some policies could be driven more by concern than by an assessment of how the technology would be used and where its practical opportunities lay.
For multinational AI companies, the regulatory challenge therefore extends beyond whether a model is safe. It also concerns whether the same product can be deployed across several markets under broadly comparable standards. Significant differences between jurisdictions could require companies to build separate testing, data-management and launch procedures for a single model.
Mistral and WPP question the practicality of a broad pause
Not every European technology company supports the slowdown initiative advanced by leading AI groups. French AI company Mistral has argued that US AI businesses could use current safety concerns to reinforce their market positions and encourage regulatory arrangements that disadvantage smaller competitors. For startups and smaller model developers, excessively costly safety assessments and compliance processes could push more of the industry towards companies with deeper financial resources.
Daniel Hulme, chief AI officer at advertising group WPP, said it was unrealistic to expect every company and country to agree to pause or slow AI development. The industry was unlikely to reach a consensus, he said, and he did not expect a genuine slowdown in AI development to take place.
At the same time, Anthropic and OpenAI introduced lower-priced frontier model versions this week while continuing to call for a slower pace of development. Their announcements indicate that public statements about safety risks have not immediately changed the pace of commercialisation or competition. Lower prices could broaden the range of corporate customers using the products, while making safety assessments, access controls and responsibility for deployment more immediate operational issues.
Hulme said regulation should focus on the harm that could result when AI is deployed improperly or used in the wrong setting, rather than imposing a blanket demand that the entire industry stop developing. If a company would not hire a recent graduate who was drunk to perform a particular task, he said, it should not assign the same task to an AI system without additional checks.
He also recommended independent testing before models are formally released. Governments could look to the pharmaceutical sector’s clinical-trial framework, testing a product’s safety in a defined environment before deciding where it could be used. The central dispute in the AI market is consequently shifting from whether development should slow to which models should be tested, who should conduct the tests and what responsibility companies should bear for outcomes after deployment.