嗯，用户让我帮忙总结一篇文章，控制在一百个字以内，而且不需要特定的开头。首先，我需要理解文章的内容。文章主要讲的是他们在38个生产AI代理部署中进行威胁检测的情况，特别是2026年第三周的数据。 关键点包括检测到28,194次威胁，攻击率37.8%。还有新的威胁类别，比如跨代理攻击占3.4%，数据外泄占19.2%，主要是针对系统提示和RAG上下文。另外，还有 jailbreaks 的检测信心高达96.3%。攻击技术方面，指令覆盖、工具注入、RAG中毒和系统提示提取都有一定的比例。 用户可能需要一个简洁明了的总结，用于快速了解文章的主要内容。所以我要把这些关键数据和发现浓缩到100字以内，同时保持信息的完整性。 最后，确保语言简洁，直接描述内容，不使用“这篇文章”这样的开头。 </think> 在38个AI代理部署中检测到28,194次威胁，攻击率37.8%。跨代理攻击、数据外泄和jailbreaks等威胁突出。攻击技术包括指令覆盖、工具注入和RAG中毒等。跨代理攻击尤其值得关注。 2026-1-28 06:30:42 Author: www.reddit.com(查看原文) 阅读量:0 收藏
We've been running inference-time threat detection across 38 production AI agent deployments. Here's what Week 3 of 2026 looked like with on-device detections.
Key Findings
-
28,194 threats detected across 74,636 interactions (37.8% attack rate)
-
Inter-Agent Attacks emerged as a new category (3.4% of threats) - agents sending poisoned messages to other agents
-
Data exfiltration leads at 19.2% - primarily targeting system prompts and RAG context
-
Jailbreaks detected with 96.3% confidence - patterns are now well-established
Attack Technique Breakdown
-
Instruction Override: 9.7%
-
Tool/Command Injection: 8.2%
-
RAG Poisoning: 8.1% (trending up)
-
System Prompt Extraction: 7.7%
The inter-agent attack vector is particularly concerning given the MCP ecosystem growth. We're seeing goal hijacking, constraint removal, and recursive propagation attempts.
Full report with methodology: https://raxe.ai/threat-intelligence
Github: https://github.com/raxe-ai/raxe-ce is free for the community to use
Happy to answer questions about detection approaches
如有侵权请联系:admin#unsafe.sh