Anthropic开放Claude Mythos 5.1网络安全访问
Anthropic给不同信任级别的安全团队分级开放Claude Mythos 5.1,GLM-5.3已被证明能生成接近Mythos的漏洞。
Anthropic通过三级访问权限向经过验证的网络防御者开放Claude Mythos 5.1,整合了Project Glasswing项目。
在CyScenarioBench攻击场景测试中,Opus 5.5每层级运行50次,无程序访问时每次都在第一个提示停止。
不同安全级别的团队获得不同权限,从分析恶意软件到测试客户系统不等。研究发现Zhipu AI的GLM-5.
3在410次尝试中构建了50个有效漏洞,接近Mythos Preview的56个。
Anthropic is opening Claude Mythos 5.1 to vetted cyber defenders through 3 access tiers that absorb Project Glasswing.
Until now, only Glasswing members got Mythos, while the Cyber Verification Program loosened filters on Opus and Sonnet for vetted security teams.
Anthropic ran Opus 5.5 through its CyScenarioBench attack scenarios 50 times per tier, and without program access every run stopped at the first prompt. anthropic
Claude now gives more hacking help the more Anthropic trusts you and the more dangerous the systems you are allowed to attack.
A hospital security team or a lone researcher gets help analyzing malware and fixing flaws, but Claude won't help them attack anything.
A penetration-testing firm can also have Claude break into client systems it is hired to test, though ransomware stays blocked.
Only organizations vetted with the US government can use Claude to probe power grids, flight systems and interbank networks.
The expansion follows Anthropic's finding that Zhipu AI's freely downloadable GLM-5.3 built working exploits in 50 of 410 attempts, nearly matching Mythos Preview's 56.
Anyone can strip that model's safety refusals, as Anthropic did for about $4,400 of compute, so attackers already have near-Mythos exploit skills and so in many sense withholding Mythos mostly handicaps defenders.
本条情报归属 AITOP「产品」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「中美大模型技术代差与商业化对照」「各国 AI 监管法案、合规治理与主权算力政策」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Claude、Mythos、Anthropic、GLM-5.3,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「rohanpaul_ai」(@rohanpaul_ai),经 AITOP 多维评分管线结构化提炼。截至目前,全网已有 11 家独立信源对该事件进行交叉印证与跟进报道(含 Anthropic、IT之家、AWS Machine Learning Blog 等),属于近期高关注度行业信号。