# 欧盟欲监管人工智能，但需OpenAI和Anthropic为监管者开门

- 来源：The Decoder：AI News（RSS）
- 作者：Matthias Bastian
- 发布时间：2026-05-12 02:19
- AIHOT 分数：55
- AIHOT 链接：https://aihot.virxact.com/items/cmp1jpods00atslve1vmpcl00
- 原文链接：https://the-decoder.com/the-eu-wants-to-regulate-ai-but-needs-openai-and-anthropic-to-let-regulators-through-the-door

## AI 摘要

OpenAI已向欧盟委员会提供直接访问其新GPT-5.5 Cyber模型进行安全审查，双方谈判已启动。Anthropic则合作困难，尽管围绕其Mythos模型举行了四到五次会议，监管机构仍无法获得访问权限。这一差距凸显了欧洲人工智能监管体系严重依赖被监管公司的自愿配合，OpenAI的开放态度与Anthropic的保留形成对比，影响了欧盟监管政策的实际执行。

## 正文

The EU wants to regulate AI but needs OpenAI and Anthropic to let regulators through the door

The EU has the AI Act and the AI Office, but still can't get hands-on access to frontier models. OpenAI's offer is welcome, but it highlights an uncomfortable truth for Brussels: EU regulators depend on AI companies volunteering access.

OpenAI has offered to give the European Commission access to its new AI model, GPT-5.5 Cyber. EU Commission spokesperson Thomas Regnier confirmed the offer during a press briefing, saying the Commission welcomes "OpenAI's transparency and their intent to give the Commission access to its new model." The access would let regulators monitor the model's deployment and address security concerns more directly.

Talks with OpenAI are already underway and will continue this week, Regnier said. Who exactly within the EU institutions will get access hasn't been decided yet. He named the EU cybersecurity agency ENISA, the AI Office, and the Directorate-General DG Connect—including its cybersecurity directorate—as potential recipients.

Regnier used the briefing to stress the importance of existing and planned EU legislation. The situation "shows once again that the EU was right to pass the legislation that we have currently in place," including the AI Act and the Cyber Resilience Act. He also pushed for getting the proposed Cybersecurity Act through the legislative process quickly.

Anthropic cooperation proves more difficult

While OpenAI approached the Commission on its own, talks with Anthropic have progressed more slowly. The Commission has already held "four to five meetings" with Anthropic about its Mythos model, the one that kicked off the current cybersecurity debate in the first place.

But the outcomes so far look different. "With one, you have a company proactively offering to give access to the company, with the other one, we have good exchanges, we're not at a stage where we can speculate on potential access or not," Regnier said.

When a Bloomberg journalist asked whether the Commission had actually requested access to Mythos, Regnier dodged the question, saying the outcome with OpenAI would be "an ideal outcome" for the Anthropic talks as well.

How such access would be secured remains an open question. Politico asked whether the Commission was preparing safeguards in case the AI Office, ENISA, or DG Connect gained access, since model access itself could become a security risk or end up in the wrong hands. Regnier didn't offer specifics, saying only that the Commission would take the necessary security precautions where needed.

The EU still can't get hands-on access to frontier AI models

European authorities currently have very little insight into Anthropic's Claude Mythos Preview model. Anthropic limited access to selected technology partners and cybersecurity companies as part of "Project Glasswing," while only the British AI Security Institute has been able to test Mythos directly.

The situation exposes a growing power gap between AI companies and European regulators, one made worse by the fact that Europe has no frontier AI companies of its own.

AI News Without the Hype – Curated by Humans