AI Security AI安全 1d ago Updated 1d ago 更新于 1天前 45

CDN Tsunami Attack Abuses HTTP/3 Translation for Up to 350x DoS Amplification CDN海啸攻击利用HTTP/3转换实现最高350倍DoS放大

Two new DoS attack vectors (HBA and HCA) exploit HTTP/3-to-HTTP/1.1 translation gaps at major CDNs, achieving up to 350x bandwidth amplification against origin servers Six CDNs tested (Alibaba, Baidu, Cloudflare, Amazon CloudFront, Fastly, Tencent) — all vulnerable to bandwidth amplification, five of six to connection exhaustion Attacks require zero configuration changes on the origin side and leverage QPACK header compression differences between HTTP/3 and HTTP/1.1 An estimated 42,330 subdomain 研究团队披露名为"CDN Tsunami"的两种DoS攻击(HBA与HCA),利用主流CDN将客户端HTTP/3流量转换为后端HTTP/1.1请求的协议转换漏洞,实现最高350倍的带宽放大效应。 攻击针对Alibaba、Baidu、Cloudflare、Amazon CloudFront、Fastly和Tencent六家CDN进行测试,所有六家均受带宽放大变体影响,五家受连接放大变体影响(Cloudflare因缓冲完整请求而免疫后者)。 核心机制在于CDN不支持端到端HTTP/3:HBA利用QPACK头部压缩将小字节请求解压为完整头部转发;HCA利用HTTP/3多路复用特性,通过低速率发送DA

65
Hot 热度
70
Quality 质量
55
Impact 影响力

Analysis 深度分析

TL;DR

  • Two new DoS attack vectors (HBA and HCA) exploit HTTP/3-to-HTTP/1.1 translation gaps at major CDNs, achieving up to 350x bandwidth amplification against origin servers
  • Six CDNs tested (Alibaba, Baidu, Cloudflare, Amazon CloudFront, Fastly, Tencent) — all vulnerable to bandwidth amplification, five of six to connection exhaustion
  • Attacks require zero configuration changes on the origin side and leverage QPACK header compression differences between HTTP/3 and HTTP/1.1
  • An estimated 42,330 subdomains across the Tranco Top 1M are potentially vulnerable, with CloudFront, Cloudflare, and Fastly having the largest exposure
  • No CVE assigned and no in-the-wild exploitation reported; mitigations are CDN-side only and include QPACK dynamic table entry size limits and reference frequency caps

Why It Matters

This research exposes a systemic architectural weakness in how CDNs translate between HTTP/3 at the edge and HTTP/1.1 at the origin — a gap that turns legitimate protocol features into powerful amplification vectors. For AI practitioners and infrastructure teams relying on CDN-hosted services, it highlights that protocol modernization at the edge without end-to-end consistency creates exploitable asymmetries with real availability consequences.

Technical Details

  • HTTP/3 Bandwidth Amplification (HBA): Exploits QPACK header compression by sending small index values that reference large headers stored in the CDN's dynamic table. The CDN must decompress these into full HTTP/1.1 headers before forwarding to the origin. Alibaba, Baidu, and Tencent achieve up to ~66x amplification using the dynamic table; Cloudflare, CloudFront, and Fastly reach 36–51x via the static table alone. Attacker bandwidth stayed below 500 Kbps–5 Mbps while origin consumption exceeded 100 Mbps.
  • HTTP/3 Connection Amplification (HCA): Leverages HTTP/3 multiplexing to open many backend TCP connections with minimal client effort. A single HTTP/3 connection carrying 96 streams can force 96 separate HTTP/1.1 connections at the origin. Five CDNs open backend connections upon receiving the HEADERS frame before the body arrives; Fastly limits this to 10 backend connections per HTTP/3 connection. Four attacker connections were sufficient to exhaust a 256-connection-limit Apache server.
  • Dynamic table constraints: Alibaba, Baidu, and Tencent advertise 4KB dynamic tables with a maximum entry size of 3,072 bytes. The amplification peaks at approximately 64 concurrent streams and then declines, attributed to CPU overhead at the CDN edge (unmeasured in the study).
  • Scope assessment: Researchers enumerated 151,685 subdomains under the Tranco Top 1M, identified 42,330 responding to HTTP/3 probes as potentially vulnerable. CloudFront led with 17,431, followed by Cloudflare (12,371) and Fastly (11,606). No origin servers outside the controlled test environment were attacked.
  • Comparison to prior work: The 2020 "CDN Judo" study found similar amplification for HTTP/2-to-HTTP/1.1 translation at ~44x (static table) and ~166x (dynamic table). HTTP/3's QPACK enables even higher factors, particularly with dynamic table exploitation.
  • Mitigations proposed: Cap QPACK dynamic table entry sizes at 512 bytes and limit references to any single dynamic table entry to no more than 10 times per stream. All mitigations are applied at the CDN layer, not the origin.

Industry Insight

  • End-to-end HTTP/3 adoption is a security imperative: The fundamental mismatch — HTTP/3 at the edge, HTTP/1.1 at the origin — is the root cause. CDNs that implement full HTTP/3-to-HTTP/3 upstream translation would eliminate both HBA and HCA attack surfaces entirely.
  • HTTP/3 default-enablement increases exposure: With Cloudflare and CloudFront enabling HTTP/3 by default for new distributions, the attack surface has grown passively. Organizations should audit their CDN configurations and advocate for origin-side protections even as edge protocol support expands.
  • Mitigation burden falls on CDNs, not origin operators: Since all proposed fixes are CDN-side and no origin-level countermeasures exist, website operators have limited recourse beyond switching providers. This creates a collective action problem where individual sites cannot fully protect themselves without vendor cooperation — reinforcing the need for industry-wide standards and potentially CVE-assigned vulnerabilities to drive urgency.

TL;DR

  • 研究团队披露名为"CDN Tsunami"的两种DoS攻击(HBA与HCA),利用主流CDN将客户端HTTP/3流量转换为后端HTTP/1.1请求的协议转换漏洞,实现最高350倍的带宽放大效应。
  • 攻击针对Alibaba、Baidu、Cloudflare、Amazon CloudFront、Fastly和Tencent六家CDN进行测试,所有六家均受带宽放大变体影响,五家受连接放大变体影响(Cloudflare因缓冲完整请求而免疫后者)。
  • 核心机制在于CDN不支持端到端HTTP/3:HBA利用QPACK头部压缩将小字节请求解压为完整头部转发;HCA利用HTTP/3多路复用特性,通过低速率发送DATA帧维持大量后端TCP连接,耗尽服务器连接容量。
  • 实测放大倍数因CDN而异:Alibaba、Baidu、Tencent支持QPACK动态表,最高达65-66倍;Cloudflare、CloudFront、Fastly最高约36-51倍;攻击者侧带宽低于500Kbps-5Mbps,而源站带宽消耗超过100Mbps。
  • 目前无CVE编号,无野外利用报告;Baidu和Tencent已确认并部署修复,但所有缓解措施均需在CDN侧实施,源站无法自行配置防御。

为什么值得看

本文揭示了HTTP/3在CDN边缘广泛部署但后端仍使用HTTP/1.1的架构缺陷,为DoS攻击提供了新的放大向量,对依赖CDN的Web服务安全具有直接警示意义。研究数据量化了不同CDN供应商的脆弱性差异,并推动了行业对协议转换层安全性的重新评估。

技术解析

  • 攻击分类与原理:HBA(带宽放大)利用QPACK动态表,攻击者发送含大头部的HTTP/3请求后,CDN将其插入动态表,后续请求仅用小型索引引用,CDN转发时需解压为完整HTTP/1.1头部,实现带宽放大;HCA(连接放大)利用HTTP/3多路复用,单个客户端连接可开启多个流,每个流触发独立后端TCP连接,通过低速率发送DATA帧保持连接活跃,耗尽源站连接池。
  • CDN表现差异:Alibaba、Baidu、Tencent支持QPACK动态表(4KB表,最大条目3072字节),带宽放大最高达66倍;Cloudflare、CloudFront、Fastly不支持动态表,放大倍数在36-51倍之间;Cloudflare因缓冲完整请求再建立后端连接,不受HCA影响。
  • 实验设置与限制:测试源站为Apache服务器,配置300秒超时和256连接限制;攻击者带宽上限30Mbps,源站上限100Mbps;放大倍数在约64并发流时达到峰值,之后因CDN边缘CPU开销而下降,但研究未提供CPU测量数据。
  • 暴露面评估:通过枚举Tranco Top 1M子域名,匹配CDN后缀并探测HTTP/3响应,发现151,685个子域名由六家CDN托管,其中42,330个可能 vulnerable,CloudFront(17,431)、Cloudflare(12,371)、Fastly(11,606)数量最多;该探测仅确认边缘支持HTTP/3,未验证实际源站是否被攻击。
  • 历史对比:与2020年"CDN Judo"研究(HTTP/2到HTTP/1.1转换)相比,HTTP/3变体放大倍数更高(静态表44倍 vs 66倍,动态表166倍 vs 66倍),表明协议演进未消除转换层风险。

行业启示

  • CDN架构安全需重新审视:协议转换层(HTTP/3到HTTP/1.1)成为新的攻击面,CDN供应商应优先实现端到端HTTP/3支持,或强化转换逻辑的速率限制与头部大小约束。
  • 源站防御能力受限:缓解措施完全依赖CDN侧实施,源站无法通过自身配置防护,这要求网站运营者在选择CDN时评估其安全响应能力,并推动行业标准制定。
  • HTTP/3部署应伴随安全评估:尽管HTTP/3提升性能,但当前CDN实现中的转换漏洞可能引入放大攻击风险,建议在启用HTTP/3时同步进行渗透测试,并关注QPACK动态表配置的最佳实践。

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。

Security 安全 Research 科学研究