Anthropic PBC 将于星期二广泛发布 Fable 5,这是一个 Mythos 版本,但会被阻止执行若干网络安全任务;数月前,该公司曾警告 Mythos 能够在关键软件中发现并利用漏洞。Claude 将把网络安全和生物学等受限制查询转由 Opus 4.8 回应,同时 Anthropic 也会向 Project Glasswing 的合资格群体发布较少防护的 Mythos 5。
Project Glasswing 上周新增 150 个组织,使具备网络能力的 Mythos 使用方总数达约 200 个。Anthropic 先前只向特定伙伴限制开放 Mythos,原因是该模型在使用者指示下可识别并利用「每个主要操作系统和每个主要网络浏览器」中的漏洞;但公司仍在推进 coding、finance 和 cybersecurity 等高价值任务能力。
Fable 5 旨在比既有模型更擅长 coding 和长时间专业问题处理。Anthropic 称,Stripe 在测试中用 1 天完成原本需团队手动 2 个月的软件工程任务;Mythos 对 E. coli 蛋白新机制提出的假设亦获研究论文确认。为验证防护栏,Anthropic 进行外部 bug bounty;逾 1,000 小时测试中,red teamers 未发现通用 jailbreak。
Anthropic PBC will widely release Fable 5 on Tuesday, a version of Mythos that is blocked from performing some cybersecurity tasks, months after warning that Mythos could find and exploit vulnerabilities in critical software. Claude will route restricted queries, including cybersecurity and biology, through Opus 4.8, while Anthropic will also release the less-guarded Mythos 5 to eligible Project Glasswing groups.
Project Glasswing added 150 organizations last week, bringing total access to the cyber-capable Mythos to about 200 groups. Anthropic had previously limited Mythos to selected partners because the model can identify and exploit vulnerabilities in “every major operating system and every major web browser” when directed by a user, while still advancing capabilities in high-value tasks such as coding, finance, and cybersecurity.
Fable 5 is intended to outperform earlier models at coding and long-duration professional problem solving. Anthropic said Stripe completed in 1 day a software-engineering task that would have taken a team 2 months manually, and a Mythos-generated hypothesis about a new E. coli protein mechanism was confirmed by a research paper. To test guardrails, Anthropic ran an external bug bounty; in over 1,000 hours, red teamers found no universal jailbreaks.