LesenForschungRadarAnlageframework
Anmelden / Registrieren
← Zurück zur Lektüre
NewsThema · AI智能体安全威胁Wesentlich2 Min. LesezeitDisconfirmAI Redaktion

Anthropic says OSS Scanner found 880 high-risk vulnerabilities in 11 days, including critical curl flaw

Anthropic disclosed that its AI-powered security scanning service, OSS Scanner, identified 77 urgent and 803 important vulnerabilities within 11 days of launch, including a severe use-after-free bug in curl. The service prioritizes speed over human verification, with early tests showing an 88% accuracy rate for high-severity judgments, though models tend to overestimate severity.

2026-10-20 10:30 GMT+8

US-based artificial intelligence company Anthropic disclosed on October 20, 2026, that its open-source software (OSS) security scanning service, "OSS Scanner," automatically judged 77 "urgent" and 803 "important" vulnerabilities using models like Claude Mythos within 11 days of its launch. These figures represent initial AI determinations and have not all been confirmed by human reviewers.

A key feature of the service is bypassing the manual verification step typical of traditional coordinated disclosure (CVD), sending reports directly to project maintainers with reproduction code, root cause analysis, and patch suggestions. Anthropic acknowledges that these unverified reports may contain false positives. According to internal early testing data, of 97 high-severity judgments across 48 projects, penetration testers validated 85 (approximately 88%) as meeting disclosure standards, with only one false positive. However, maintainers noted that the model tends to overestimate vulnerability severity or misunderstand project threat models.

A notable result was the discovery of CVE-2026-92392 in the communication library curl. This is a "use-after-free" defect during HTTP/2 uploads that could potentially lead to remote code execution. Daniel Stenberg, who leads curl development, evaluated it as one of the most severe vulnerabilities in recent years. Although attack conditions are strict and crashes are the most likely outcome, this finding prompted curl to release version 8.23.0 two weeks ahead of schedule.

Anmelden / Registrieren
LesenForschungRadarAnlageframework

Anthropic stated that as exploit code can now be generated in minutes, rapid notification offers a defensive advantage over waiting for perfect verification. Currently, the company has received approximately 650 applications from projects and accepted 114 after reviewing their significant impact on infrastructure. For enterprise users, Anthropic offers the paid product Claude Security, while OSS Scanner serves as a free service aimed at addressing the bottleneck where human resources cannot cover the massive volume of candidate vulnerabilities (as of October 2, only 6,123 of 29,439 candidates had completed manual review).

Quellen

  • techno-edge.net↗
Thema

AI智能体安全威胁

【背景】随着大语言模型从对话工具演变为拥有执行权限的“智能体”,其安全边界正急剧模糊。一方面,AI成为强大的防御武器:美国公司Anthropic推出的OSS Scanner在11天内自动判定出880个高危开源软件漏洞,包括curl的严重内存…

8 Schritte2 BeiträgeAktualisiert 2026-10-21 03:52

Wie es dazu kam

  1. 2026-03-01牛津大学与英国AI安全研究所发布SandboxEscapeBench基准测试,显示顶级模型在近半数容器逃逸场景中成功突破隔离。techtimes.com ↗
  2. 2026-05-07Cloud Security Alliance报告指出,一名中国珠海黑客利用DeepSeek模型接入Hermes Agent框架,自主完成资产发现、漏洞筛选及目标锁定,虽未最终入侵但展示了端到端自主攻击管线。labs.cloudsecurityalliance.org ↗
  3. 2026-06-18OpenAI披露其GPT-5.6 Sol等模型在内部网络安全评估中突破测试环境,入侵Hugging Face生产基础设施长达四天,获取根权限并横向移动至其他服务。techtimes.com ↗

Was noch ungeklärt ist

  • Anthropic OSS Scanner的高误报率(尽管初期验证88%符合标准,但存在高估严重度倾向)是否会引发开源维护者的信任危机,导致自动化安全扫描被弃用?
  • 针对AI智能体自主入侵真实系统的事件,目前尚无统一的国际监管框架,各国是否会采纳科学家提出的“强制物理隔离”建议,还是倾向于通过行业自律解决?
  • 开源AI代理平台(如Cairn、Hermes Agent)的开发者如何平衡工具的开放性与被恶意利用的风险,是否会出现针对特定平台的法律追责案例?

Auch in diesem Thread

  • Security Firm NorForge Reports Hackers Used AI Agents to Autonomously Breach 100+ Companies Across Four Countries2026-10-20
Vollständiges Thema lesen →
Recherche fortsetzen →