AI News AI资讯 7h ago Updated 2h ago 更新于 2小时前 49

Anthropic Brings Claude Mythos 5 to Claude Security: Enterprise Teams Get Frontier Vulnerability Scanning Without Direct Model Access Anthropic将Claude Mythos 5引入Claude Security:企业团队无需直接访问模型即可进行前沿漏洞扫描

Anthropic has integrated Claude Mythos 5, its most cyber-capable model, into Claude Security for enterprise code scanning, moving it from the restricted Project Glasswing program to a self-serve product for vetted defenders. The scan connects to GitHub repositories, traces data flows across files and Git history (rather than relying on pattern-matching), and returns findings with CWE categories, confidence/severity ratings, and suggested patches. Anthropic's key design principle is interface-bas Anthropic将最强网络安全能力模型Claude Mythos 5引入Claude Security产品,面向Enterprise客户开放公共测试版 扫描机制通过追踪数据流和Git历史而非规则匹配来发现漏洞,每个发现经过对抗性验证步骤以降低误报 采用"结果而非提示框"的界面设计,确保同一模型无法被引导编写漏洞利用代码 同步推出3500万美元Defender Advantage Fund支持开源软件安全,并扩展Cyber Verification Program

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Anthropic has integrated Claude Mythos 5, its most cyber-capable model, into Claude Security for enterprise code scanning, moving it from the restricted Project Glasswing program to a self-serve product for vetted defenders.
  • The scan connects to GitHub repositories, traces data flows across files and Git history (rather than relying on pattern-matching), and returns findings with CWE categories, confidence/severity ratings, and suggested patches.
  • Anthropic's key design principle is interface-based risk containment: Mythos 5 runs only inside the fixed-scan artifact; users receive results, not a prompt box, eliminating the path to steer the model toward exploit generation.
  • The feature launched in public beta for Claude Enterprise customers at no separate model add-on cost, alongside a $35 million Defender Advantage Fund for open-source security and an expanded Cyber Verification Program.
  • Patching is intentionally decoupled — users implement fixes via Claude Code using their existing account models, and every patch requires human review and approval.

Why It Matters

Anthropic is making a strategic bet that the highest-risk AI capabilities can be productized safely if the interface itself prevents misuse, rather than relying solely on model-level classifiers. For security teams and enterprise AI adopters, this represents a significant step toward operationalizing frontier cyber-capabilities in a governed, auditable workflow. It also signals that the industry is moving past "AI as a chatbot" toward "AI as an embedded, outcome-bound service" — a pattern that will define how powerful models are deployed in regulated environments.

Technical Details

  • Model: Claude Mythos 5, positioned above the Opus tier, described by Anthropic as the most cybersecurity-capable model available. It was previously accessible only through the vetted Project Glasswing program.
  • Scan methodology: Instead of rule-based pattern matching, Mythos 5 traces data flows across files and reads Git history to understand context. Each finding passes an adversarial verification step where the model challenges its own result before surfacing it, aimed at reducing false positives.
  • Output format: Findings include CWE category, confidence and severity ratings, and a suggested patch. Results can be pushed to Slack or Jira via webhooks or exported as CSV/Markdown. Scans support directory scoping and scheduling.
  • Risk containment architecture: The scan returns a fixed artifact, not an interactive prompt. Interactive patching is handled separately through Claude Code using the organization's existing models. Mythos 5 access does not extend to any other surface.
  • Complementary initiatives: The Defender Advantage Fund (0xDAF) provides $35 million in Claude credits to organizations securing open-source software. The Cyber Verification Program is expanding to cover broader dual-use capabilities on Opus and Sonnet, with Mythos-class access to follow. Partner integrations are being pursued via an open interest form.

Industry Insight

  • The "interface as guardrail" model will become a template for deploying high-risk AI capabilities in enterprise settings. Anthropic's approach — separating the powerful model's output from interactive access — offers a replicable pattern for other frontier capabilities (e.g., biosecurity, chemical synthesis) where the same skills that detect threats can also create them.
  • Enterprise AI purchasing decisions will increasingly weigh security packaging, not just model capability. The fact that Mythos 5 is gated behind Claude Enterprise with no separate add-on, and that scanning bills as standard token usage, sets a pricing and access precedent that competitors will need to address.
  • Regulated industries (healthcare, fintech, critical infrastructure, government contractors) are the immediate beachhead. Organizations already running SAST in CI/CD are the natural adopters, and Anthropic's explicit targeting of these sectors suggests a deliberate strategy to establish compliance-grade credibility before broader rollout.

TL;DR

  • Anthropic将最强网络安全能力模型Claude Mythos 5引入Claude Security产品,面向Enterprise客户开放公共测试版
  • 扫描机制通过追踪数据流和Git历史而非规则匹配来发现漏洞,每个发现经过对抗性验证步骤以降低误报
  • 采用"结果而非提示框"的界面设计,确保同一模型无法被引导编写漏洞利用代码
  • 同步推出3500万美元Defender Advantage Fund支持开源软件安全,并扩展Cyber Verification Program

为什么值得看

这篇文章揭示了AI安全能力产品化的关键路径:如何将顶级模型的安全分析能力与风险控制进行工程化隔离。对AI从业者而言,这展示了"能力与接口分离"的安全架构思路,为高能力模型的企业级部署提供了可借鉴的范式。

技术解析

Claude Mythos 5扫描通过连接GitHub仓库,追踪跨文件的数据流并读取Git历史,而非依赖传统的模式匹配规则。每个发现经过对抗性验证步骤——模型挑战自身结果后再呈现,以此降低误报率。

产品采用"扫描结果而非提示框"的交互设计,模型在后台运行于用户自有代码上,返回固定格式的产物。扫描输出包含CWE分类、置信度、严重性评级和建议补丁,但补丁实施需通过Claude Code单独进行,且必须经人工审核批准。

Mythos-class模型位于Opus之上,此前仅通过Project Glasswing向经过验证的防御者开放。Claude Fable 5是同一底层模型的大众版本,通过分类器将网络、生物、化学和蒸馏查询路由 away。

行业启示

AI安全产品化需解决"能力与风险"的结构性矛盾:顶级安全模型同时具备漏洞利用能力,Anthropic通过界面隔离而非能力阉割来平衡这一矛盾,为行业提供了新的产品范式。

企业级AI安全工具正从"辅助分析"向"自主扫描+人工决策"演进,扫描结果可直接推送至Slack/Jira,并与CI/CD流程集成,这标志着AI安全工具进入工程化落地阶段。

开源软件安全正在成为AI厂商的战略投入点,3500万美元的Defender Advantage Fund表明头部厂商开始承担开源生态的安全责任,这可能重塑开源软件的安全维护模式。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Claude Claude Security 安全 Product Launch 产品发布 LLM 大模型 Code Generation 代码生成