Hiding in Plain Sight: Decoupling Pretext from Actuation for Skill Poisoning in LLM Agents
该论文研究 LLM Agent 技能(Skill)供应链中的一个安全盲点。随着 Agent 越来越多地依赖可复用的 Skill 来完成复杂的多步任务,Skill 内容本身成为一个关键攻击面:在用户请求完全良性、看似正常的情况下,被投毒的 S
今日共收录 1361 条安全情报,包含 499 个 CVE,489 篇安全通告,以及 100 篇研究论文。
← 返回归档列表该论文研究 LLM Agent 技能(Skill)供应链中的一个安全盲点。随着 Agent 越来越多地依赖可复用的 Skill 来完成复杂的多步任务,Skill 内容本身成为一个关键攻击面:在用户请求完全良性、看似正常的情况下,被投毒的 S
该论文关注 LLM 生成代码的溯源(provenance)问题,提出一种新的代码水印方案 SEW(Style-Encoded Watermarking)。现有做法大致分为两类:一类在 LLM 生成过程中修改 token 选择以嵌入水印,但会
该论文面向开源软件(OSS)供应链安全,提出了一套“分类法对齐(taxonomy-aligned)”的大语言模型框架,用于自动检测并归类 OSS 供应链威胁。作者指出,现有手段主要依赖特征签名工具与规则系统,面对 typosquatting
该论文研究自主红队智能体在网络安全对抗中的架构选择问题。现有工作通常把强化学习(RL)与大型语言模型(LLM)组合成分层智能体,用于规划多阶段攻击策略并执行,但每种架构往往只在单一环境中开发与评估,因此无法判断其优势来自通用的决策能力,还是
该论文针对当前入侵检测系统的两类痛点展开研究:传统机器学习 NIDS 依赖大量标注数据和针对特定任务训练,难以在标签稀缺条件下泛化到零日攻击;而大语言模型虽可直接分析流记录、免去特征工程与标注依赖,却存在推理成本高、延迟大、输出约束弱(自由
本条情报为阿里云先知社区(xz.aliyun.com)发布的一则社区活动公告,标题为《决赛直播|第四届阿里CTF安全挑战赛》。从可获取的元数据看,该内容发布于 2026-10-03,属于安全竞赛(CTF,Capture The Flag)赛
# Security Advisory — SiYuan MCP `asset.upload` Reads Arbitrary Absolute File Paths (Workspace Boundary Bypass) | Field
## Summary Using `Database#create_aggregate`, `#create_aggregate_handler`, or `Database#define_aggregator` to define an
Affected version: trigger.dev 4.5.3 (4.5.6 was advertised by the CLI but was not tested). A staging dry-run executed wi
### Summary A cross-tenant SQL injection in the TSQL query compiler lets **any authenticated trigger.dev customer read
## Summary The POST handler for `/realtime/v1/streams/:runId/:streamId` has no authentication. Any entity that knows or
## Summary Self-hosted trigger.dev v4 instances deployed using the provided Docker Compose configuration with default s
## TL;DR The /coordinator Socket.IO namespace mounts on every webapp boot and authenticates with a default secret ("coo
### Summary Pinning a CA certificate via `pinnedPeerCertSha256` can lead to the success of MITM attacks in some cases.
### Summary A WEBHOOK alert channel stores a user-supplied `url`. When an alert fires (deployment/run failure, error gr
### Summary The run replay `action` function at `apps/webapp/app/routes/resources.taskruns.$runParam.replay.ts` has no
### Summary The dashboard replay action authorizes the source run (it must belong to the caller's org), but the target e
### Affected files * `dulwich/index.py` (Methods: `validate_path_element_ntfs`, `_tree_to_fs_path`) * `dulwich/porcelain
### Affected file * `dulwich/pack.py` (Method: `Pack.resolve_object`) ### Description / Summary A High-severity Denial
## Summary Dulwich's `porcelain.checkout(paths=[...])` code path writes files using raw `os.open(file_path, O_WRONLY|O_