Anthropic CEO Dario Amodei 近日行程罕见低调：他本周原定在牛津郡会见50位来自欧洲大型公司的高管，并在英国停留仅1到2天。公司同时把 Mythos 描述为过于危险而未公开发布的模型，却已让“数十个”机构试用预览，显示其在谨慎与扩张之间的拉扯。

Anthropic 一方面为 Mythos 的市场化和可能的IPO做准备，另一方面又把自己包装成负责任的社会角色。它与瑞士的 Financial Stability Board 进行闭门简报，讨论网络风险；本周还派联合创始人 Christopher Olah 参与 Pope Leo 的AI议题，同时公司却在起诉 Pentagon，争议焦点是其技术在战争中的使用。

在伦敦面向开发者的活动上，数百名来自金融、能源和初创企业的工程师聚集，关注 Claude Code。会上展示了更强的自主性：数十甚至数百个AI代理可在无需提示的情况下连续运行数小时；一位大型华尔街银行开发者担心，系统越独立，出事后越难追踪责任链。Anthropic 的产品负责人 Cat Wu 则称系统“incredibly secure”，问题更多在于公司如何解释其安全设计。

Anthropic CEO Dario Amodei has recently been unusually hard to pin down: he was due to meet 50 executives from major European companies in Oxfordshire this week and stay in the UK for only 1 to 2 days. At the same time, the company is presenting Mythos as a model too dangerous to release, even as it has allowed “a few dozen” organizations to test and preview it, underscoring the tension between caution and expansion.

Anthropic is preparing Mythos for the market and possibly for an IPO, while also casting itself as a responsible social actor. It has held closed-door briefings with the Financial Stability Board in Switzerland about cyber risks, and this week sent co-founder Christopher Olah to join Pope Leo on AI, even as the company is suing the Pentagon over how its technology may be used in war.

At a developer event in London, hundreds of engineers from finance, energy and startups gathered around Claude Code. Demonstrations showed greater autonomy: tens or hundreds of AI agents can run tasks continuously for hours without prompting. One Wall Street bank developer said the more independence AI tools gain, the harder it becomes to trace accountability if something goes wrong, while product head Cat Wu said the system is “incredibly secure” and that the larger issue is explaining its security design.