2019年，OpenAI完成训练新模型GPT-2后，最初宣布其过于危险而暂缓发布，但同年后又公开发布。七年后，Anthropic 的现任负责人Dario Amodei 在4月7日表示，最新发布的Claude家族模型 Mythos 目前仍不应“广泛可用”。

Anthropic 声称 Mythos 的能力“远超此前训练过的任何模型”，并指出其可发现软件漏洞并用于修复或攻击。该公司报告该模型已在“每个主要操作系统和浏览器”中发现严重漏洞，其中包括一个隐匿27年的漏洞，同时其年化营收从去年底的90亿美元上升到300亿美元，增长了约2.33倍。

Anthropic 还同步推出 Glasswing 计划，与Apple、Linux Foundation、CrowdStrike及Google合作，让企业在发布前用 Mythos 测试未公开代码；公司先支付前1亿美元使用成本，但后续费用将是前代模型 Opus 的5倍。该举措加剧了行业趋势：其他公司也将推出具备类似攻防能力的模型，而美国对“零日漏洞”的战略储备与国家安全考量可能因此受到挑战。

In 2019, after OpenAI completed training GPT-2, it initially said the model was too dangerous to release but later released it later that year. Seven years on, Dario Amodei, now at Anthropic, said on April 7 that the newly released Claude-family model Mythos is still not ready for wide release.

Anthropic says Mythos is substantially more capable than any model it has trained before and can detect software vulnerabilities that can be either fixed or exploited. It says Mythos has found severe flaws in every major operating system and browser, including one that went undetected for 27 years, while Anthropic’s annualised revenue reportedly rose from $9bn at the end of last year to $30bn on April 6, a roughly 3.33x increase.

Anthropic paired the warning with Project Glasswing, involving firms like Apple, the Linux Foundation, CrowdStrike, and Google, so companies can test unpublished code with Mythos before release; Anthropic covers the first $100m of initiative costs, then charges five times Opus pricing. This reflects a broader trend: rivals are likely to launch models with similar cyber-offense capabilities, while U.S. concerns over hoarded zero-days and broader security policy, including pressure from the defense establishment, may intensify if Mythos exposure lowers the secrecy value of hidden vulnerabilities.

Source: How dangerous is Mythos, Anthropic’s new AI model?

Subtitle: Dario Amodei’s warnings should not be dismissed

Dateline: 4月 09, 2026 03:38 上午