EU opens talks with OpenAI, Anthropic

The European Commission confirmed on May 11, 2026 that it has opened strategic discussions with OpenAI and Anthropic to better understand the behaviour and governance of large, general‑purpose AI models ahead of upcoming enforcement milestones under the EU AI Act.

Brussels framed the meetings as an effort to give regulators direct visibility into “frontier” model capabilities and to clarify compliance needs for systems that could pose systemic risks. The engagement is being described as a technical and governance dialogue rather than a commercial negotiation.

OpenAI has stepped forward with an offer to give EU institutions and selected European partners access to a cybersecurity‑focused variant of its latest model, which Brussels welcomed as a constructive step for monitoring deployments and mitigating harm. Company announcements and Commission briefings indicate the offer covers access for EU cyber authorities and the European AI Office.

By contrast, Anthropic has held several meetings with Commission officials but — according to Brussels — those exchanges have not yet moved to system‑access arrangements. EU spokespeople said talks with Anthropic were productive but not at a stage where the Commission could confirm access to the firm’s advanced models.

The focus on Anthropic stems from concerns around its Mythos preview, a model that several security and finance bodies say can surface novel vulnerabilities and — in testing — produced complex exploit‑like behaviours. Regulators and some national finance ministers have urged access to such tools so European banks and supervisors can test and harden defenses.

The timing matters because the AI Act’s supervisory architecture is rolling forward in stages. The European AI Office and other enforcement tools have been phased in since 2025, with key GPAI and high‑risk obligations tied to an August 2026 enforcement milestone that will give the Commission broader powers to investigate and compel information from general‑purpose AI providers.

That staged timetable is one reason Brussels wants early, technical contact with labs. Under the AI Act, providers of general‑purpose AI models face documentation, red‑teaming, incident reporting and transparency obligations — requirements that are easier to test when regulators can inspect model specifications and observe behaviour in controlled settings.

Member states have added political pressure. Several finance ministries and central banks urged Brussels to secure access to advanced models so critical infrastructure and banks can be stress‑tested, and some national officials have suggested the EU should seek reciprocal testing arrangements with providers. Those requests have pushed the Commission to move from rule writing toward direct engagement with labs.

Commission officials say access arrangements could take different forms: supervised previews, data‑safe testbeds, formalised cooperation with the AI Office or contractual access for market surveillance. The EU also expects firms to implement the General‑Purpose AI Code of Practice and to provide the technical documentation the law requires.

Industry reaction is mixed. Some security teams in Europe welcomed OpenAI’s offer as pragmatic, while other experts warned that selective access could entrench dependence on non‑EU providers and complicate rules about extraterritorial compliance and market surveillance. Analysts say the talks highlight a broader debate about Europe’s technological sovereignty.

The European Parliament is taking the topic to the chamber level: a May 18 plenary session in Strasbourg lists a briefing on EU readiness for emerging, potentially cyber‑capable AI models and the interplay between the AI Act, NIS2 and the Cyber Resilience Act. Lawmakers will press the Commission on how access arrangements will protect citizens and critical infrastructure.

For now, Brussels describes the dialogue as preparatory — an intelligence‑gathering and risk‑management exercise designed to ensure model deployments in Europe meet the Act’s obligations and that national supervisors can do their work. With the AI Office’s full enforcement powers approaching, the stakes for both labs and EU regulators are rapidly sharpening.