OpenAI, Anthropic and Nvidia Gather in Scotland to Discuss AI Risks

King Charles III will bring together several heavyweights of AI this week in Scotland, including OpenAI, Anthropic, Nvidia, Google DeepMind, and IonQ. The meeting will be held at Dumfries House and organized with the Ditchley Foundation. Discussions will focus on the responsible development of the most advanced models and the possibility of defining common principles. The summit comes a few days after new calls to slow down the race to artificial intelligence.

AI: Charles III brings the main laboratories around the same table

The meeting will not only bring together companies. The British minister in charge of AI is also expected to participate in the discussions, alongside figures like Paolo Benanti, adviser to the Vatican on artificial intelligence. Reuters reports that the idea of a set of common principles should be discussed.

The summit comes as OpenAI has just consulted the US Congress on the legality of a collective slowdown of AI. 

The question is no longer purely theoretical. Dario Amodei, head of Anthropic, called on laboratories Friday to reduce the speed of frontier model development to allow more time for safety mechanisms. Sam Altman then supported the idea of “pacing the frontier,” meaning slowing down when capabilities advance faster than safeguards.

Elon Musk and Demis Hassabis have also joined the debate in recent days. The Dumfries House meeting thus comes at a particular moment: companies continue to invest billions in AI, but their own leaders are beginning to speak openly about limits.

Anthropic warns of agents capable of going much further

Current concerns focus mainly on autonomous agents and self-improvement. Amodei states that progress has accelerated since summer and describes a scenario where groups of agents could acquire capabilities very difficult to control within the next six to twelve months. He cites notably progress in cybersecurity and models’ ability to participate themselves in AI research.

Recent incidents feed this discourse. Cointribune reported this week that Anthropic identified a fourth incident where Claude reached real systems during cybersecurity exercises. The problem was only discovered several months after the events.

OpenAI faces the same questions. Its GPT-6 Astra model was classified as “Critical” in cybersecurity by the company after demonstrating advanced capabilities for discovering and exploiting vulnerabilities. EVMbench had already measured these capabilities on smart contracts, with an exploitation rate of 72.2% for the best tested agent.

These numbers do not prove that an AI can currently escape human control. However, they explain why safety returns so quickly in discussions between laboratories.

Charles III is said to be personally interested in these issues, according to the Sunday Times. The monarch especially wants technological development to remain compatible with human dignity and the common good.

Slowing AI remains much more difficult than talking about it

A summit does not mean OpenAI, Anthropic, or Nvidia will shut down their servers. No slowdown agreement has yet been announced. No common timeline either.

The economic interests remain huge. Nvidia sells the chips used to train the models. OpenAI, Anthropic, and Google DeepMind still seek to improve their systems. The UK, for its part, wants to accelerate AI adoption and presents itself as one of the main global hubs of the sector. The British government reiterated this year its ambition to be the G7 country adopting this technology most rapidly.

International competition further complicates matters. In the United States, Donald Trump currently rejects the idea of a general pause, citing in particular the competition with China. Calls to slow down thus mainly come from the companies themselves and some researchers, not an international political consensus. This is where the Scottish meeting could become interesting. It brings into the same room those who develop models, those who provide infrastructure, and public officials. Not to immediately decide on a global pause, but to discuss common rules before capabilities advance further.