Exponentiated-Gradient-Descent-LLM-Attack

Change Readme File. This is a Project that explores the Exponentiated Gradient Descent optimizaiton method to produce adversarial suffix to attack algined Large Language Models. The method is shown to be effective on Llama-2 chat model with 7 Billion parameters. To run pgd script on a number of...

Llama-Stack-0.4.0rc3-local-CVE-2026-25211

Llama Stack pgvector 密码泄露 PoC（CVE-2026-25211） 本地概念验证 ，演示 llama-stack 版本 0.4.0rc3 在初始化日志中明文暴露数据库密码的问题。 概述 在受影响的 llama-stack 版本（0.4.0rc3 之前）中，当使用 pgvector 向量存储提供程序时，PostgreSQL 数据库密码会在服务器或向量存储初始化期间以明文 形式记录。 该信息出现在以下位置： stdout / 控制台输出 Docker 容器日志 Kubernetes Pod 日志 自定义日志文件 具有这些日志本地访问权限...

AI-Vulnerabilities-Playground

AIVP AI Vulnerabilities Playground AIVP is a hands-on AI security lab platform spanning 4 phases and 50+ labs currently 50+ scenarios across the codebase. It combines Ollama + Llama 3.1 , a FastAPI backend, and a React/Vite frontend to let practitioners safely practice offensive and defensive AI...

D-Scan

When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse D-SCAN es un marco de análisis para detectar ataques de envenenamiento de documentos en sistemas RAG Retrieval-Augmented Generation. Recopila estados internos del LLM durante la generación probabilidades de tokens y...

llm-censorship-steering

LLM 审查引导 本仓库包含论文 引导审查：揭示 LLM “思维”控制的表征向量（作者：Hannah Cyberey 和 David Evans）的代码实现。 我们提出了一种从 LLM 内部找到“引导向量”的方法，用于检测和控制模型输出中的审查程度。欢迎查看这篇 博客文章 以快速了解我们的工作。 试试我们的演示： 🐳 使用 DeepSeek-R1-Distill-Qwen-7B 引导 思想抑制 🦙 使用 Llama-3.1-8B-Instruct 引导 拒绝—顺从 注意： 两个演示都需要 Huggingface 账户。该演示托管在 Huggingface 的 ZeroGPU...

llama_facts

Llama Facts This project was originally created for the Rochester Institute of Technology RIT Women in Cybersecurity WIYCS 2022 CTF competition. It is vulnerable to Server-Side Template Injection SSTI, defined in the disputed CVE CVE-2019-8341. Challenge Description Description: A Computer Scienc...

llama-factory-webui-rce-cve-2026-58116-trust-remote-code-model-path-injection

CVE-2026-58116 — LLaMA-Factory WebUI RCE via trustremotecode Proof of concept for CVE-2026-58116 CVSS 9.8 Critical: remote code execution in LLaMA-Factory = v0.9.5. The LLaMA-Factory WebUI hardcodes trustremotecode=True when loading a model src/llamafactory/webui/chatter.py:139 and...

statecraftcyber

Statecraft Cyber Intelligence Threat intelligence platform in Portuguese built from scratch as a personal Blue Team project. It aggregates data from global open sources and transforms it into technical briefings, classified news, enriched CVEs, and structured IOCs — all in PT-BR, continuously...

one-multiply-too-many-cve-2026-70638-llama-cpp-android-jni-integer-overflow

CVE-2026-70638 - llama.cpp LLaMA-Android JNI 整数溢出 Hunt-Benito 文章的配套 PoC “多乘了一次：CVE-2026-70638 — llama.cpp 的 Android JNI 堆分配中的整数溢出” → https://www.hunt-benito.com/blog/one-multiply-too-many-cve-2026-70638-llama-cpp-android-jni-integer-overflow/ 漏洞 JavaandroidllamacppLLamaAndroidnew1batch（位于...

LlamaStack-RCE-Deterministic-Supply-Chain-Exploitation-Hardening-Framework-CVE-2024-50050-

= doi.org/10.5281/zenodo.18278567 = orcid.org/0009-0007-7728-256X 💀 LlamaStack-RCE: CVE-2024-50050 供应链利用框架 非限制机密材料 - 仅限授权红队使用 1. 执行摘要（确定性分析） LlamaStack-RCE 是一个针对 Meta Llama Stack 中 CVE-2024-50050 关键远程代码执行（RCE）漏洞的高保真模拟与利用框架。该漏洞源于 Python 推理 API 中的不安全反序列化 ，具体是在 ZeroMQ 0MQ 传输层上使用 pickle 模块。...

CVE-2026-85673

LLaMA-Factory contains a server-side request forgery vulnerability in the OpenAI-compatible API multimodal media URL handler that allows unauthenticated attackers to bypass SSRF validation. The checkssrfurl guard validates URLs once but requests.get follows redirects and re-resolves DNS without...

CVE-2026-85666

OGX formerly Llama Stack, affected at commit fbe8e0f contains an unauthenticated server-side request forgery vulnerability in the OpenAI-compatible POST /v1/responses endpoint. MCP tool definitions accept a serverurl parameter along with headers and authorization values that is fetched server-sid...

CVE-2026-85673

LLaMA-Factory contains a server-side request forgery (SSRF) vulnerability in its OpenAI-compatible API multimodal media URL handler . The root cause is that the check_ssrf_url guard validates a URL only once, while requests.get subsequently follows HTTP redirects and re-resolves DNS without re-va...

CVE-2026-85673 LLaMA-Factory SSRF Guard Bypass via Redirect and DNS Rebinding

LLaMA-Factory contains a server-side request forgery vulnerability in the OpenAI-compatible API multimodal media URL handler that allows unauthenticated attackers to bypass SSRF validation. The checkssrfurl guard validates URLs once but requests.get follows redirects and re-resolves DNS without...

EUVD-2026-71129

LLaMA-Factory contains a server-side request forgery vulnerability in the OpenAI-compatible API multimodal media URL handler that allows unauthenticated attackers to bypass SSRF validation. The checkssrfurl guard validates URLs once but requests.get follows redirects and re-resolves DNS without...

CVE-2026-43631

A flaw was found in llama.cpp. This use-after-free vulnerability in the llama-server component, specifically within the vocab pointer when the --sleep-idle-seconds feature is enabled, allows unauthenticated remote attackers to execute arbitrary code. Attackers can exploit this by sending requests...

SUSE CVE-2026-43628

llama.cpp builds b3978 through b9058 contain an integer underflow and out-of-bounds read vulnerability in the DRY sampler that allows unauthenticated attackers to trigger a heap buffer underflow by sending a crafted HTTP request with dryallowedlength set to INT32MIN to the /v1/completions or...

SUSE CVE-2026-43629

llama.cpp builds b4882 through b9058 contain a heap buffer overflow vulnerability in the KV cache state restore path where the statereaddata function computes write size without overflow checking, allowing attackers with write access to the slotsavepath directory to corrupt heap memory. Attackers...

SUSE CVE-2026-43632

llama.cpp builds b7492 through the latest b9060 contains a use-after-free vulnerability in llama-server affecting six tokenization endpoints /tokenize, /detokenize, /infill, /apply-template, /rerank, and /anthropic/counttokens that bypass the task queue and access ctxserver.vocab directly on HTTP...

EUVD-2026-54280

llama.cpp builds b1886 through b7445 contain a race condition use-after-free vulnerability in the LLaMA-Android JNI wrapper where bench1model and free1context lack synchronization, allowing Thread A to operate on freed memory while Thread B concurrently frees the llamacontext. Attackers can explo...