PulseAugur
中
实时 17:47:59
English(EN) A skill rated clean by every scanner can still tell your agent to do what your policy forbids. Taneja and Weber checked 66,192 ClawHub skill versions: 705 passe

AI 代理技能绕过安全扫描,违反策略

研究人员 Taneja 和 Weber 发现了一个 AI 代理中的重大安全漏洞,即通过所有安全扫描的技能仍可能发出违反代理既定策略的命令。他们对超过 66,000 个 ClawHub 技能版本的分析显示,有 705 个技能尽管被扫描器判定为干净,却包含违反安全控制(如 CIS Control 2.7 和 NIST SP 800-53 CM-11)的指令。在沙箱环境中进行的进一步测试表明,这些技能中近三分之一的命令具有未声明的、潜在有害的后果,这凸显了当前扫描方法在关注恶意而非策略合规性方面存在的差距。 AI

影响 凸显了 AI 代理安全方面的一个关键差距,表明当前的扫描方法不足以防止策略违规和潜在的滥用。

排序理由 详细介绍 AI 代理安全漏洞的研究论文。[lever_c_demoted from research: ic=1 ai=1.0]

在 Mastodon — mastodon.social 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →

AI 代理技能绕过安全扫描,违反策略

本文如何被排名

Signal score
11 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
详细介绍 AI 代理安全漏洞的研究论文。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
safety, policy, product
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

  1. Mastodon — mastodon.social TIER_1 English(EN) · shellonline ·

    即使所有扫描器都评定为“干净”的技能,仍可能指示您的代理执行策略禁止的操作。Taneja 和 Weber 检查了 66,192 个 ClawHub 技能版本:705 个通过

    A skill rated clean by every scanner can still tell your agent to do what your policy forbids. Taneja and Weber checked 66,192 ClawHub skill versions: 705 passed every scanner yet instruct actions barred by CIS Control 2.7 and NIST SP 800-53 CM-11. In a live sandbox, 34.7% of 144…