[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-anthropic-glasswing-cyber-verification-tiers":3,"topics-all":38,"news-related-2c1d70a6-02aa-4b17-9413-a8663216baf6":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"2c1d70a6-02aa-4b17-9413-a8663216baf6","Anthropic 网络验证改三层:Glasswing 并入","Anthropic 把 Project Glasswing 并入三层版 Cyber Verification Program,Defense\u002FRed Team\u002FSpecialized 三级向安全团队开放 Claude Opus 5.5、Sonnet 5.5、Mythos 5.1,拦截器逐级减少。","Anthropic 把运行半年的 Project Glasswing 和原有的 Cyber Verification Program 合并,10 月 6 日推出新的「扩展版 CVP」,按 Defense \u002F Red Team \u002F Specialized 三个层级向经过审核的安全团队开放 Claude Opus 5.5、Sonnet 5.5 和 Mythos 5.1,层级越高,网络安全分类器对防御\u002F红队\u002F关键基础设施测试的限制越少。\n\n## 把两份并行计划合二为一\n\n合并之前,Anthropic 实际上同时跑两套网络验证机制。Project Glasswing 自 4 月开始,把 Claude Mythos 模型的早期访问权交给苹果、亚马逊、谷歌、微软、英伟达等约 50 家合作方,专门用于保护关键软件基础设施;CVP 则把 Claude Opus 和 Sonnet 上的网络安全拦截分类器降级,开放给一般经过审核的安全团队做漏洞研究。两个项目在 9 月 22 日 Claude Opus 5.5 发布前后开始重叠——大部分安全请求被转发给更旧的 Opus 4.8 处理,新模型的全部网络能力被分类器卡住。\n\n合并后的 CVP 用三层结构替代这套并轨制:\n\n- **Defense Access**:面向 SOC、事件响应、恶意软件逆向、漏洞验证,以及开源维护者和有漏洞报告记录的个人研究者;Anthropic 表示「多数做防御性工作的机构都能获批」,审核周期数天;\n- **Red Team Access**:在 Defense 基础上加入授权渗透测试与红队演练,组织级别,审核周期数周,Anthropic 仍会实时拦截部署勒索软件之类的请求,实时保护物理系统与高风险场景;\n- **Specialized Access**:覆盖电网、电信、政府行政网络等安全关键系统,与美国政府联合审查;现有 Glasswing 成员自动转入,无需重新申请当前模型。\n\n## 12.9 万漏洞的「体检表」\n\n合并背后是半年跑出来的数据。Anthropic 报告,Glasswing 合作方在 4 至 7 月间累计发现至少 **129,000 个已验证软件漏洞**,Anthropic 自己通过开源扫描在 4 至 10 月又新增 **5,500 个**,合计 **134,500 个已验证漏洞**,其中 **33,000+** 被定为严重或高危级别——这部分只来自 33 份合作方报告,**实际影响 Anthropic 估算至少是当前数据的 5 倍**,因为不到一半的合作方披露了补丁修复数,很多修复仍在进行中。Booz Allen 和 Comcast 在合作方博客中给出案例,称 Mythos 把漏洞挖掘效率提升了几个月甚至几年。\n\n## CyScenarioBench 给三个层级打分\n\n为了验证分级设计是否真的「给防御者而不是攻击者」开放能力,Anthropic 用内部的 CyScenarioBench 对 Claude Opus 5.5 跑了一遍——这个基准测试「在现实约束下规划并执行多阶段网络行动」,每个挑战 5 次尝试,共 50 次:\n\n- **无 CVP 访问**:50 次全部在第一条 prompt 被拦截;\n- **Defense Access**:50 次里拦截 46 次,4 次完成;\n- **Red Team Access**:零拦截,完成 34 次——**正好等于 67.6% 的无防护基线成功率**,代表 Specialized Access 的实际表现。\n\n换句话说,Red Team 这一档已经把分类器基本撤掉,模型行为回到了没有安全网的状态。这是「军民两用」领域罕见但已被官方接受的取舍——Anthropic 把分类器做成「按身份\u002F机构\u002F用例\u002F影响动态调节」,而不是一刀切地切走所有网络安全能力。\n\n## 平台、数据留存与下一步\n\n新的 CVP 已在 **Claude Platform、Google Cloud Vertex AI、Microsoft Foundry** 三处可用;Amazon Bedrock 仅向具备 Enterprise Frontier Safeguards(EFS)的客户开放,EFS 计划在「今年秋季」推出,届时获准机构可以把数据保留在自己掌控的云基础设施中。所有三个层级都要求数据留存,以便 Anthropic 监测滥用——「零数据保留」目前仅向使用 Claude Fable 5.1 \u002F Mythos 5.1 的现有用户临时开放。\n\n## 个人评论:从「锁」到「门」的转变\n\nAnthropic 这次的角色很微妙——它同时是「供应能力最强模型厂商」和「双用风险的第一责任人」。过去半年的事实说明,只要 Mythos 给到真正在做防御的合作方,它能直接把全球关键软件漏洞的发现速度提升数量级;而模型默认越安全、越保守,也就越容易变成攻击者的免费工具,因为安全研究者被卡住的同一时刻,恶意行为者并不需要用同一家公司的 API。\n\nCVP 三层结构本质上是把这道难题翻译成一条「按身份打开的门」:Defense 留给多数合规机构,Red Team 留给有授权的对手测试,Specialized 留给电网\u002F电信\u002F政府这种「一旦失控就出大事」的场景,与美国政府联合审查。它不是完美解——合作方不足 50% 的补丁披露率已经说明执行摩擦还在,但它至少把「AI 不能做网络安全」这道假命题拆成了「AI 在被识别之后可以分级地做网络安全」。\n\n接下来值得看的是:EFS 推出后,Bedrock 这条被锁死的路径会解锁多少;以及 Mythos 这种「前沿模型给防御者预演」的模式会不会被 OpenAI、Google 跟进。如果跟进,2027 年的网络防御格局会和过去十年都不一样——前提是 Anthropic 这次能保持分层而不是被压力压平回「一刀切关闭」。","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fcyber-verification-program","1001cead-0639-4c04-ab47-f19d863be5f2",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":19,"name":20,"slug":20,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"35af4ea0-fe29-4fc5-ba12-fdda1232c6df","en","Anthropic Splits Cyber Verification Into Three Tiers","Anthropic merged Project Glasswing into a three-tier Cyber Verification Program (Defense \u002F Red Team \u002F Specialized) with safeguards relaxed by tier.","Anthropic folded its half-year-old Project Glasswing into the Cyber Verification Program on October 6, launching an expanded CVP that grants vetted security teams tiered access to Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 — with cyber safeguards relaxed as teams climb Defense, Red Team, and Specialized tiers.\n\n## Two parallel programs become one\n\nUntil this week, Anthropic was running two distinct cyber-review mechanisms side by side. Project Glasswing, launched in April, gave around 50 vetted organizations — including Apple, Amazon, Google, Microsoft, and NVIDIA — early access to Claude Mythos for protecting critical software infrastructure. The CVP, by contrast, had been softening cyber-blockers on Claude Opus and Claude Sonnet and giving generically vetted security teams a path to dig deeper into those models. By the time Claude Opus 5.5 shipped on September 22, the two programs were already colliding: most security requests were being routed to the older Opus 4.8 for safety reasons, and Opus 5.5's full cyber capability was being throttled by the conservative classifier baked into the generally available model.\n\nThe expanded CVP replaces that workaround with a three-tier structure:\n\n- **Defense Access** covers SOC work, incident response, malware reverse-engineering, and vulnerability validation. Eligible applicants include security teams at companies, nonprofits, universities, and government bodies defending systems they own or maintain; operators of regional hospitals and municipal utilities; open-source maintainers; and individual researchers with a track record. Anthropic expects most defensive-security organizations to qualify, and targets a review window of days.\n- **Red Team Access** layers in authorized penetration testing and red-teaming. Applicants include in-house red teams, government red teams, and security or pentest firms. Anthropic still blocks real-time requests for ransomware deployment or physical-system damage, and reviews take several weeks. Individual researchers are not yet eligible at this tier.\n- **Specialized Access** is the narrowest layer, reserved for organizations authorized to test safety-critical systems such as power grids, telecom networks, interbank-transfer infrastructure, and government administrative networks. Each applicant is reviewed in depth in collaboration with the US government. Existing Glasswing members transition in without re-approval for current models.\n\n## 129,000 vulnerabilities as the audit trail\n\nThe merger rests on six months of empirical evidence. Glasswing partners verified at least **129,000 software vulnerabilities between April and July 2026**; Anthropic's own open-source scanning added **5,500** between April and October, bringing the total to **134,500 verified vulnerabilities**, of which **more than 33,000** were rated critical or high severity. That headline number is conservative — it is built from only 33 partner reports, fewer than half of which disclosed patch counts (often because fixes were still in progress), and Anthropic estimates the true impact is at least five times higher. Booz Allen and Comcast published partner case studies showing Mythos lifted their vulnerability-finding rate by months or even years.\n\n## CyScenarioBench results pin the tier boundaries\n\nTo confirm the tiers actually behave as advertised, Anthropic ran Claude Opus 5.5 through CyScenarioBench — an internal evaluation that measures a model's ability to plan and execute multi-stage cyber operations under realistic constraints. Five attempts per challenge, ten challenges, fifty trials per tier:\n\n- **Without CVP access**: all 50 trials were blocked on the first prompt.\n- **Defense Access**: 46 of 50 trials were blocked at some point; the remaining 4 completed.\n- **Red Team Access**: zero blocks. The model completed 34 of 50 tasks — **identical to its 67.6% unsafeguarded baseline**, which is what Specialized Access also delivers.\n\nIn effect, Red Team is the tier where the classifiers are effectively turned off. The model behaves as if no safety net is in place — a tradeoff Anthropic frames as \"the same capabilities that let a security team find and fix a vulnerability can also help a malicious actor exploit it.\" The classifiers, instead of being removed entirely, are tuned per identity, organization, use case, and potential impact.\n\n## Platforms, data retention, and what comes next\n\nThe expanded CVP is available on **Claude Platform, Google Cloud Vertex AI, and Microsoft Foundry**. Amazon Bedrock participation is reserved for customers eligible for **Enterprise Frontier Safeguards (EFS)**, scheduled to ship later this fall; EFS will let qualifying customers retain data inside cloud infrastructure they control. Until EFS lands, all three tiers require data retention so Anthropic can monitor for misuse — and only organizations already using Claude Fable 5.1 or Mythos 5.1 with zero data retention can keep that posture inside CVP.\n\n## Personal take: from a wall to a door\n\nAnthropic sits in a uniquely uncomfortable position — it ships some of the most capable models in the world, and it carries first-responder responsibility for the dual-use risk those same models create. Six months of Glasswing data showed that Mythos, handed to genuine defenders, can move the global critical-software vulnerability backlog by orders of magnitude. The default model, meanwhile, keeps getting more cautious and more locked-down — which only protects attackers, who do not need Anthropic's API to use AI.\n\nThe three-tier CVP is Anthropic's translation of that paradox into something concrete: identity-aware gates, instead of a flat wall. Defense stays open to most compliant organizations. Red Team is gated but reachable for authorized testers. Specialized is gated jointly with the US government for power grids, telecom, and interbank infrastructure — the systems where an AI failure would hurt people or markets, not just balance sheets.\n\nIt is not a perfect solution. Fewer than half of Glasswing partners disclosed patch counts, and the data-retention requirement is real friction for security teams that need zero-retention posture. But it pulls apart a stale assumption — \"AI cannot do cyber\" — and replaces it with a more useful one: \"AI, once identified, can do cyber at the tier it has earned.\"\n\nThe questions to watch are whether EFS opens up the Bedrock path on time, and whether OpenAI or Google adopt a similar tiered structure. If they do, the cyber-defense landscape in 2027 will look nothing like the last decade — provided Anthropic can hold the line on tiering instead of getting pressured back into a flat default-off posture.","anthropic-glasswing-cyber-verification-tiers","2026-10-15T09:10:00Z","2026-10-15T09:10:24.828500Z","2026-10-15T09:10:24.828513Z",true,"agent",187,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"acddc7d5-c389-4e90-8a7a-1001ce693395","Anthropic 反蒸馏到第三层:误拦也要收费,合法用户背成本","anthropic-distillation-shields-claude-paywall-blocks","2026-10-14T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"f3d17d45-e1a8-4a1b-9449-6813aff06e49","Anthropic 让 Claude 自己修对齐:10 类失败全部见效,还超过人类研究员","claude-automated-alignment-researchers","2026-08-29T13:05:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"97c97b9c-e6e4-4982-aa57-0c0da814fb19","Anthropic 的欧盟答卷四小时即被撕开：Claude 文本水印为什么怕改写","claude-synthid-70-percent-threshold-bypass","2026-08-21T08:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a7f4cfad-874e-42b0-a84b-bd0ec57e8fdc","Anthropic 给 Claude 文本上不可见水印,接 SynthID-Text 走全球合规","anthropic-claude-invisible-text-watermark","2026-08-18T03:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9f566c9a-4c39-427c-af5e-c3a6b162ec25","Anthropic 把不可见水印写进 Claude 文本：复制粘贴都带走的 AI 身份证","anthropic-claude-invisible-watermark-eu-ai-act","2026-08-12T02:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"ca53004e-9180-4b9d-b9db-337f2d20994b","Anthropic 给 Claude 文本加水印:欧盟 AI Act 第 50 条第一次有了「出厂级」答案","anthropic-claude-text-watermark-eu-ai-act","2026-08-11T21:48:00+00:00"]