Inside Elastic InfoSec's agentic SOC: How we cut AI agent LLM calls by 60%
We run fourteen AI agents that triage Elastic InfoSec alerts. They were taking 19 LLM calls to do work that needed 8. He
#detection
共收录 84 条相关情报。
← 返回所有主题Silence False Alarms: Identifying Anti-Reentrancy Patterns on Ethereum to Refine Smart Contract Reentrancy Detection.
以太坊智能合约的重入漏洞曾导致巨大的经济损失，为此社区开发了多种自动化重入检测器。然而，这些检测器因检测规则过于粗糙而频繁产生大量误报，常将已采用防重入模式（如互斥锁、检查-生效-交互模式等）保护的合约误判为存在漏洞。现有代码分析技术在识别
Inside Elastic InfoSec's agentic SOC: When to inline your agent's skills for a 5× cost reduction
We tested two agentic SOC architectures in parallel across 36,822 real Agent Builder conversations. One won by 5.7x: a s
How Elasticsearch ES|QL COMPLETION turns noisy curl and wget rules into high-fidelity cloud security alerts
Elastic InfoSec tested this detection rule pattern on their own cloud fleet, filtering noisy curl and wget events with d
wp2shell hits WordPress: detecting pre-auth RCE from plugin drop to command execution
We ran the wp2shell WordPress RCE chain end-to-end with Elastic Defend. Detection rule walkthrough, IOCs, and hunt guida
Deepfake News Detection: A Multimodal Framework Integrating LipNet, DeepSpeech and ResNET for Enhanced Audio-Visual Analysis
本文提出了一种多模态深度伪造新闻检测框架，旨在通过同时分析视频中的音频和视觉线索来鉴别视频内容的真实性。该框架首先从唇部运动（使用LipNet编码）、音频内容（使用DeepSpeech2编码）和视频帧（通过BlazeFace检测人脸并用Re
NetRadar: Enabling Robust Carpet Bombing DDoS Detection.
地毯式轰炸攻击（Carpet Bombing Attack）是一种日益流行的分布式拒绝服务（DDoS）变体，它同时向受害网络中的多台服务器发起攻击，令每条恶意流的流量速率极低，以绕过传统基于流量阈值的检测机制。聚合后的恶意流量会淹没网络接入
On the Detectability of Active Gradient Inversion Attacks in Federated Learning.
联邦学习（FL）允许多个客户端协同训练机器学习模型，同时将数据保留在本地，从而提升隐私保护。然而，FL训练过程中交换的梯度仍易受到梯度反转攻击（GIA）的威胁，攻击者可以重建客户端本地数据，破坏FL的隐私承诺。GIA可由被动或主动的服务器发
New North Korean campaign uses fake coding interviews to steal developer credentials
DPRK-aligned hackers hid malware inside SVG flag images to backdoor developer job interview coding tests. Not one antivi
TELEPUZ: a modular MaaS malware spreading via CLICKFIX-VIDAR chains
TELEPUZ is a modular malware that emerged through CLICKFIX-VIDAR attacks in April. We reverse-engineered it to show you
PVDetector: Detecting Prompt Injection Attacks on Purpose-Specific LLM Agents through Policy-Violation Concept Analysis
随着大型语言模型（LLM）被越来越多地部署为针对特定目的智能代理（如客服、代码生成），它们不仅要遵守通用安全护栏，还需遵循针对其角色设计的特定目的限制。这些额外限制扩大了攻击面，特别是提示注入（PI）攻击。现有的检测方法主要依赖输入-输出模
ClickFix to Cash-Out: Anatomy of a Mexican Banking-Fraud Toolkit
Elastic Security Labs tracks REF6045, an active operator-assisted banking fraud operation targeting customers of Mexican
Generalized Synthetic Image Detection with Enhanced RGB-Noise Representation Learning
本文针对大规模生成模型快速发展导致的高度逼真AI生成图像泛滥问题，提出了一种新的通用合成图像检测框架RNSIDNet。现有取证网络通常依赖单一域表示和传统二分类优化，导致跨模型泛化能力差且对真实世界退化（如JPEG压缩、缩放等）鲁棒性不足。
A Penny for Your Prompts: Experiments Detecting and Mitigating LLM Usage by Survey Respondents
本研究针对众包平台（如Prolific和Mechanical Turk）上调查问卷参与者越来越多使用大型语言模型（LLM）辅助回答的现象，旨在量化这种行为的普遍性并探索检测与缓解方法。研究团队开展了一系列调查（样本量N=250），系统评估了
Tracing Target Answers in Poisoned Retrieval Corpora via Token Influence Attribution
该论文针对检索增强生成（RAG）系统面临的语料库投毒攻击问题，提出了一种轻量级检测框架TRACE。RAG系统通过检索外部文档来增强大语言模型的生成能力，但攻击者可向检索库中注入恶意文档，诱导模型输出特定目标答案。现有检测方法通常依赖额外的分
Lost in relocation: analysis of a new loader distributing CASTLESTEALER
Find out how a new obfuscated loader evades static detection using .reloc section abuse, five anti-VM/language checks an
Azure AD Graph Activity Logs: Ingestion and threat detection to close the visibility gap
Azure AD Graph Activity Logs land in Elastic with full ECS parsing. Detect ROADrecon and AADInternals enumeration with r
MGTBench: Benchmarking Machine-Generated Text Detection.
本文提出了 MGTBench，一个用于评估机器生成文本 (MGT) 检测方法的统一基准测试框架。随着以 ChatGPT 为代表的大型语言模型 (LLM) 在文本分类、情感分析、翻译、问答等 NLP 任务中展现出强大能力，机器生成文本的检测变
Cheap Reward Hacking Detection
该研究提出一种低成本、高效的reward hacking检测方法，针对强化学习（特别是基于人类反馈的强化学习，RLHF）中奖励模型被操纵的问题。核心方法：训练一个小型Transformer编码器，将Terminal-Wrench环境中的轨迹
An Embarrassingly Simple Detector for Model Extraction Attacks in Large Language Model API Traffic
本研究针对大型语言模型（LLM）通过托管API部署时面临的模型提取攻击威胁。模型提取攻击中，攻击者通过发送大量查询来窃取或复制目标模型的功能，但单个查询往往与正常用户请求难以区分。现有检测方法多基于单条查询异常评分或纯良性用户与攻击者用户分