405 matches found
hcaptcha-challenger
文档 快速入门 简介 hCaptcha Challenger v0.13.0+ 借助大语言模型的 Spatial Chain-of-Thought SCOT 能力构建了一套端到端的 Agentic Workflow，使得 Agent 可遵循指令完成空间视觉通用任务，而无需额外训练或微调 CNN 专家模型。 在 hCaptcha Challenger 中，Agent 通过 Playwright 控制浏览器页面。在您的任务流程中，Agent 通过传入的 Page 对象初始化，并借用该对象接管当前页面的交互。您可以通过 Agent 实现 clickcheckbox（点击验证框）和...
bordair-multimodal
Multimodal Prompt Injection Dataset 516,588 个带标签样本 （251,782 个攻击样本 + 251,576 个良性样本，外加一个 13,230 样本的真实世界验证划分），涵盖五个数据集版本及外部数据集摄入，覆盖跨模态、多轮、对抗性后缀、越狱模板、间接注入、工具操纵、智能体、规避、推理 DoS、视频生成、VLA 机器人、LoRA 供应链、原生音频 LLM、RAG 优化、MCP 跨服务器、编码智能体、序列化边界以及智能体技能供应链等针对 AI 系统的攻击。攻击样本与良性样本按 1:1 平衡（审计清理后比率为 0.9992:1）。...
Multimodal-Unlearnable-Examples
多模态不可学习样本：保护数据免受多模态对比学习的影响 ACM MM2024 摘要...
ai-captcha-bypass
AI 驱动的 CAPTCHA 求解器 该项目是一个基于 Python 的命令行工具，利用大型多模态模型（LMM），如 OpenAI 的 GPT-4o 和 Google 的 Gemini，来自动解决各种类型的 CAPTCHA。它借助 Selenium 实现浏览器自动化，与网页交互并实时解决 CAPTCHA。 成功求解的动图会以 GIF 格式保存在 successfulsolves 目录中。 详细博客文章 ：https://aydinnyunus.github.io/2025/12/08/ai-captcha-bypass/ 主要特性 多 AI 提供商 ：支持 OpenAI（例如...
vLLM 0.8.3 - 0.14.0 - Information Disclosure
vLLM 0.8.3 to - 0.14.1 contains an information disclosure caused by leaking a heap address in error messages from the multimodal endpoint when processing invalid images, letting remote attackers reduce ASLR entropy, exploit requires sending invalid images. id: CVE-2026-22778 info: name: vLLM 0.8....
Multimodal-web-attack-Dataset
MWAD: Un conjunto de datos multimodal de ataques web para la detección de inyección SQL basada en IA Descripción general El conjunto de datos multimodal de ataques web MWAD es un novedoso conjunto de datos etiquetado de ciberseguridad desarrollado como parte de la investigación del autor en...
EXPLOIT-CVE-2026-22778
CVE-2026-22778 — vLLM 视频处理中的 RCE 针对 CVE-2026-22778 （CVSS 9.8）的漏洞实验环境与概念验证，该漏洞是 vLLM 多模态摄取路径中的未认证远程代码执行链。 | ---|--- CVE| CVE-2026-22778 公告| GHSA-4r2x-xpjr-7cvv 受影响版本| vLLM = 0.8.3, vLLM 将媒体加载失败转换为 HTTP 400，并将 exc.detail 原封不动地返回给客户端 （apiserver.py）： async def httpexceptionhandler: Request, exc:...
safety-awareness
跨模态安全漂移的迁移安全意识 跨模态安全漂移中多模态大语言模型的迁移安全意识 （EMNLP 2026 Findings）的官方实现。 本仓库包含 Qwen3-VL 安全意识方向流水线。它从有害和无害多模态示例中提取与拒答相关的激活方向，选择能够改善拒答行为同时限制良性任务漂移的方向，并可选地训练一个方向特定的安全意识向量。 当前代码通过 Hugging Face Transformers 支持 Qwen/Qwen3-VL-8B-Instruct。 仓库结构 . ├── artifacts/directions/ 预计算的选定方向及元数据 ├── data/csv/ 流水线使用的 CSV ...
Learning-to-Detect
Learning to Detect Unknown Jailbreak Attacks in Large Vision-Language Models Official implementation of “Learning to Detect Unknown Jailbreak Attacks in Large Vision-Language Models.” This repository contains the data-processing, hidden-state extraction, classifier training, safety-pattern...
CVE-2026-103243
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
CVE-2026-103243
Versions of LightLLM up to and including 1.2.0 are vulnerable to a Server-Side Request Forgery (SSRF) . The root cause is a failure to validate the image_url and audio_url parameters within multimodal endpoints . This allows unauthenticated attackers to supply arbitrary URLs to fetch internal res...
CVE-2026-103243 LightLLM through 1.2.0 Server-Side Request Forgery via multimodal endpoints
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
EUVD-2026-89988
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
CVE-2026-103243 LightLLM through 1.2.0 Server-Side Request Forgery via multimodal endpoints
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
CVE-2026-103243 LightLLM through 1.2.0 Server-Side Request Forgery via multimodal endpoints
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
CVE-2026-103243: Server-Side Request Forgery (SSRF)
LightLLM through 1.2.0 fails to validate imageurl and audiourl parameters in multimodal endpoints, allowing unauthenticated attackers to perform server-side request forgery. Attackers can supply arbitrary URLs to fetch internal resources, with vision model processing disclosing content or error...
CVE-2026-103041
LightLLM through 1.2.0 multimodal deployments expose an unauthenticated RPyC cache service with pickle deserialization enabled on all interfaces. Attackers can send crafted serialized objects to exposed cache methods to execute arbitrary code with service privileges...
CVE-2026-103041
LightLLM versions through 1.2.0 in multimodal deployments are vulnerable to Remote Code Execution (RCE) . The flaw exists because an unauthenticated RPyC cache service is exposed on all interfaces with pickle deserialization enabled. An attacker can send crafted serialized objects to the exposed ...
CVE-2026-103041 LightLLM through 1.2.0 Unauthenticated Remote Code Execution via Embed Cache RPyC Service
LightLLM through 1.2.0 multimodal deployments expose an unauthenticated RPyC cache service with pickle deserialization enabled on all interfaces. Attackers can send crafted serialized objects to exposed cache methods to execute arbitrary code with service privileges...
CVE-2026-103041 LightLLM through 1.2.0 Unauthenticated Remote Code Execution via Embed Cache RPyC Service
LightLLM through 1.2.0 multimodal deployments expose an unauthenticated RPyC cache service with pickle deserialization enabled on all interfaces. Attackers can send crafted serialized objects to exposed cache methods to execute arbitrary code with service privileges...