《WIRED》报导指出，Anthropic 在 2026 年 4 月 7 日正式宣布 Claude Mythos Preview，同时推出跨产业的 Project Glasswing，以应对新模型带来的网路安全冲击。该联盟包含 Microsoft、Apple、Google、Amazon Web Services、Linux Foundation、Cisco、Nvidia、Broadcom 以及超过 40 家其他科技、资安、关键基础设施与金融组织；另有资料指涉超过 45 家成员。Anthropic 将 Mythos Preview 先行提供给成员企业私密使用，并未全面发布，目的是让各大平台开发者先在自有系统中测试，降低漏洞被即时滥用的风险。
Anthropic 这次行动的核心讯号不只在某一家公司，而是警示「未来 6、12、24 个月」内 AI 能力将普及，传统资安假设可能失效。Dario Amodei 表示，Mythos 并未专门训练成为网安工具，但因其程式码能力突出，副作用是产生强大的网路攻击与防御能力。Logan Graham 指出，Mythos Preview 不仅能找出漏洞并产生攻击链与概念验证，还能进行高级攻击利用开发、渗透测试、端点安全评估、系统错误设定侦测，以及在无法存取原始码下评估软体二进位。
Anthropic 采取阶段式发布，并参照协同揭露漏洞流程：先让少量受信任组织先知道并修补，再公开讨论。Logan Graham 甚至称其能力已可达到资深资安研究员的水准，若未妥善释出，可能成为攻击者的放大器。Google、Microsoft 等成员公开支持：Heather Adkins 指 AI 同时带来资安挑战与机会，Igor Tsyganskiy 也表示已藉 Mythos Preview 发现「成千上万」重点漏洞，甚至包含被忽略多年的旧漏洞。Graham 强调，这不应只剩少数公司实验的专案，而必须持续扩大，否则会失败。
WIRED reports that Anthropic officially announced Claude Mythos Preview on April 7, 2026, while also launching Project Glasswing to address cybersecurity implications. The consortium includes Microsoft, Apple, Google, Amazon Web Services, the Linux Foundation, Cisco, Nvidia, Broadcom and over 40 other organizations in technology, cybersecurity, critical infrastructure, and finance, with reporting also describing more than 45 participants overall. Anthropic is giving members private access to Mythos Preview rather than a general release, so platform developers can test first on their own systems and reduce the window where newly discovered flaws can be exploited.
The strategic message is broader than one vendor: Logan Graham says teams must prepare for capabilities broadly available in 6, 12, and 24 months, because current security assumptions may no longer hold. Dario Amodei said Mythos was not trained specifically for cybersecurity but became strong in cyber by being strong at code. Graham says Mythos Preview can generate vulnerabilities and exploit chains, produce proof-of-concept attacks, and also do advanced exploit development, penetration testing, endpoint security assessment, misconfiguration hunting, and source-free binary assessment.
Anthropic is using a staged release model inspired by coordinated vulnerability disclosure: selected members get early access, then results can be fixed before broad discussion. Graham says Mythos can match what a senior security researcher can do; without careful release, it could accelerate attackers significantly. Launch statements from partners were notably collaborative: Heather Adkins of Google framed AI as both a new threat and an opportunity, while Igor Tsyganskiy of Microsoft said the preview had already exposed thousands of critical vulnerabilities, including longstanding bugs. Graham reiterated that Glasswing must become much larger than a handful of firms to avoid failure.