Project Glasswing
Project Glasswing is Anthropic's early-access program for AI models, designed to test and validate model capabilities and safety in collaboration with selected partner organizations. As part of a staged deployment strategy, the program collects real-world feedback before broad release to evaluate model performance boundaries and identify potential risks, forming one component of Anthropic's responsible deployment framework.
Timeline (last 90 days)
Anthropic has merged Project Glasswing and its original verification program into a unified Cyber Verification Program, divided into three tiers—Defender, Red Team, and Special Access—covering Claude Opus 5.5, Sonnet 5.5, and subsequent new models.
Anthropic shared Mythos's vulnerability-discovery capabilities with a small group of selected companies to help them audit source code and patch vulnerabilities, in a program named 'Project Glasswing'.
Some Reddit users speculated that the biggest outcome of Project Glasswing may have been to allow Anthropic to train its subsequent model Fable on closed-source code that it otherwise would not have been able to access.
All Facts (5)
Project Glasswing是Anthropic推出的前沿AI模型早期访问计划
90%UnverifiedAnthropic的Mythos Preview（Project Glasswing）是一个利用大语言模型进行自主漏洞发现的内部研究项目
60%UnverifiedAnthropic shared Mythos's vulnerability-discovery capabilities with a small group of selected companies to help them audit source code and patch vulnerabilities, in a program named 'Project Glasswing'.
50%UnverifiedSome Reddit users speculated that the biggest outcome of Project Glasswing may have been to allow Anthropic to train its subsequent model Fable on closed-source code that it otherwise would not have been able to access.
50%UnverifiedAnthropic has merged Project Glasswing and its original verification program into a unified Cyber Verification Program, divided into three tiers—Defender, Red Team, and Special Access—covering Claude Opus 5.5, Sonnet 5.5, and subsequent new models.
50%