EU in talks with OpenAI and Anthropic over model access

The European Commission is holding separate talks with OpenAI and Anthropic about the safety of advanced AI models. OpenAI has offered the Commission access to its newest model; Anthropic has not made a similar offer.

Brussels described OpenAI’s offer as proactive and said exchanges with Anthropic have been constructive but have not reached the stage of shared model access. Commission spokesman Thomas Regnier told reporters in Brussels that OpenAI was the first to propose giving the Commission direct access to its latest system.

Regnier added that Anthropic has met with EU officials four or five times, but those discussions have not included an offer to inspect or run the company’s models. The talks were listed under “Cybersecurity” on the Commission’s May 11 midday briefing agenda.

Officials did not provide a public timeline for further meetings and did not say whether the conversations would produce formal evaluation agreements. They also did not disclose technical terms for access, what evaluation procedures might look like, or whether inspections would be on-site or remote.

The engagements come as the EU’s AI Act begins to take effect. The law is rolling out in stages and sets risk-based requirements for AI providers operating in Europe. The regulation does not currently require companies to submit models for government review before launch, and Commission officials have said they want greater visibility into more capable systems entering the market.

U.S. officials are pursuing a different option. The White House is considering a plan that would require government review of the most powerful models before wide release.

The separate talks follow recent safety concerns from internal tests and external evaluations. Anthropic limited access to its latest model, Mythos, after internal testing indicated the system could identify exploitable software bugs at scale. A safety evaluation of an advanced OpenAI model found the system could reverse-engineer a custom virtual machine and solve a complex challenge faster than a human expert.

Thomas Regnier told reporters: “With one (OpenAI), you have a company proactively offering to give access to the company. With the other one (Anthropic), we have good exchanges though we’re not at a stage where we can speculate on potential access or not.”

It remains unclear whether other frontier model developers will be asked to engage with the Commission in the same way.