adr (@aaaddress1) header
adr (@aaaddress1) profile picture

Team Lead of PSIRT and Threat Research, TXOne Networks. Speaker at Black Hat USA, CODE BLUE, DEFCON, HITB, HITCON, S4, SECTOR. Author of Windows APT Warfare

台灣
Joined December 2013
Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)Photo by adr (@aaaddress1)
adr retweeted
AI 做网络安全,过去两年大家都在卷谁家 Agent 挖 0day 更快。但今天 Google DeepMind 发布的 AlphaProof Code 展示了另一条足以改写系统软件工程历史的路线:直接让多 Agent 军团把真实生产环境的工业级加密库,用数学证明焊死到机器码层级。 懂形式化验证(Formal Verification)的人都知道这四个字有多贵:当年为了数学证明 seL4 微内核,顶尖学者整整砸了 20 个人年;CompCert 编译器砸了 6 个人年,而且前提是源码本身就得为了配合证明专门重写。 而 DeepMind 这次拿来开刀的靶子,是全球数十亿人加密通信的底座——Signal 官方生产库 libsignal: 1. 不改一行源码,直接证到 AArch64 机器码 没有专门改写源码迁就证明器,而是直接拉取 未修改的 Rust 生产代码(libsignal + spqr),用标准 rustc 编译成真实的 AArch64 二进制可执行文件,加载进 Lean 的硬件指令集语义里做端到端精化证明,连编译器优化可能引入的偏差一并兜底。 2. 2.6 万行受审规范撬动 650 万行零信任机器证明 AI 军团一共肝出了约 650 万行 Lean 证明代码,但人类不需要去读这 650 万行——安全根基只建立在人类专家严格审定的约 26,000 行 规范与硬件语义上。无论 Prover Agent 怎么搜索试错,最终必须同时过两道铁闸:Lean 内核校验 + 独立 Comparator 核对未篡改的隔离规范副本。 3. 拿下后量子握手与三重棘轮核心安全定理 历史上第一次把协议层的前向保密、认证性、后妥协自愈安全(PCS)与握手保密,一路无断层地连通到了广泛部署的真实密码学库编译产物上。 只要规范写对,AI 生成形式化证明天生不存在幻觉漏网之鱼——因为 Lean 内核是一个铁面无私的确定性裁判。以前只有航天和微内核才烧得起的形式化验证,终于要被 AI 打成常规武器了。
5
16
1
126
7,897
AI 圈已经有人开始为「AI 断供」做末日准备了,甚至囤出了 10.5TB 的 AI 战略储备😂 黄仁勋离开 NVIDIA 后,Ruby on Rails 创始人 DHH 突然发了一条意味深长的推文: 「你最好提前备份几份开放权重模型,再准备一些能运行它们的基本设备。没人知道 NVIDIA 到底发生了什么,但这个世界刚刚失去了自由开放 AI 最重要的支持者之一。」 Reddit上,已经有人认真讨论起了「如果有一天开放权重模型不再发布,我们该提前囤些什么?」 结果评论区直接变成了 AI 末日生存物资展示大会。 一位叫 Kahvana 的玩家晒出了自己的 2TB NVMe 硬盘。 里面不只是几个聊天模型,而是囤了 Qwen3.8、Gemma 4、DeepSeek V4 Flash、Nemotron 等多个系列,连不同精度的量化版本都准备好了。 而且考虑得相当周全: 不仅有负责聊天和写代码的大语言模型,还有语音识别、语音合成、图片生成、Embedding、Reranker,以及 ComfyUI、MCP 工具、离线维基百科。 甚至还专门把最重要的模型复制到 NAS,并备份 llama.cpp 和 SillyTavern,确保不只是有模型权重,还能真正运行起来。 但这还不是最夸张的。 另一位叫 TheBigHODL 的玩家直接晒出了一份 10.5TB 的 AI 储备清单。 中国的 Kimi、DeepSeek、Qwen、GLM、MiniMax,美国的 Llama、GPT-OSS、Gemma,以及 Mistral 等模型,几乎把主流开放权重阵营都囤了一遍。 还不止如此。 从 Whisper 语音模型、FLUX 图片模型,到训练数据集、评测数据、完整维基百科、推理框架、量化工具,甚至一些开发工具的源码,都在他的存档范围内。 这已经不是简单地下载几个 GGUF 文件了,而是在试图给自己保存一整套能独立运转的 AI 工具箱。 他们担心的也很现实: 万一未来中国模型受到更多出口或访问限制,美国 AI 实验室又逐渐停止开放权重,今天随手就能从 Hugging Face 下载的东西,明天可能就没那么容易拿到了。 当然,这只是社区对未来的担忧,并不代表相关禁令已经出台。 但这种变化确实挺有意思。 以前本地 AI 圈比的是谁的显卡多、显存大、Token 跑得快。 现在有人已经开始比谁的硬盘大、谁的模型囤得全、谁断网之后还能继续用 AI。 以前末日生存玩家囤粮食、柴油、发电机,现在 AI 玩家开始囤模型、算力和离线知识库了。 而且仔细想想,这事还真不完全是杞人忧天。 云端 AI 再强,使用权终究取决于服务商;本地模型即使落后几个版本,只要权重、软件和硬件都在,至少还能自己运行。
128
442
68
3,194
366,067
I’ll be presenting new vulnerabilities uncovered during my DKnife research at @BlackHatEvents Europe! See you in London in a few weeks!
Black Hat Europe Speaker Spotlight Series 🎬 Meet Ashley Shen, @ashl3y_shen , Security Research Engineering Technical Leader, Cisco Talos, in our latest Speaker Spotlight as she answers three key questions: 👉 What are you most excited about at Black Hat Europe? 👉 What will your session focus on? 👉 What’s one key takeaway attendees can expect? In Black Hat Europe Briefing, “Eyes on the Edge: From an In-the-Wild AitM Framework to an IP-Camera Vulnerability Chain,” she shares research spanning active adversary-in-the-middle (AitM) campaigns and a real-world vulnerability chain affecting internet-connected IP cameras. Learn more: bit.ly/47tmWQc
2
24
929
adr retweeted
[VXCON 2026] Celebrating 16 Years of Hacking: Deep-Dive into Exclusive Offensive Security & AI Tracks (Register Now!) Dear Cybersecurity Professional, The wait is over. VXCON is returning for its landmark 10th Anniversary edition this November 28–29, 2026, in Hong Kong. Since our journey began in 2010, VXCON has dedicated the last 16 years to building an elite community of white-hat hackers, malware analysts, reverse engineers, and offensive security researchers. To celebrate this milestone, we are bringing together global trail-blazers who frequently speak at Black Hat, DEF CON, and HITCON. They will break down live vulnerabilities, exploit pathways, and hardware weaknesses you won’t see anywhere else. Secure your seat today to learn from the best in the industry: vxcon.hk Featured Keynotes & Technical Briefings Our 2026 lineup features exclusive research spanning OS kernels, automated AI hacking agents, automotive systems, and critical infrastructure. 🛡️ Windows Kernel & Browser Exploitation • Angelboy (MSRC Top 100) covers cross-component vulnerabilities in the Windows Ancillary Function Driver (AFD), detailing transport-layer gadgets and sandbox escapes via Zoom. • MinJeong Kim explores variant hunting in WebKit by tracing patches to root causes and introduces the AI-driven WebKit Weekly framework. 🤖 The Frontier of AI & Automation in Security • Wonyoung Jung debuts a hybrid LLM and program analysis framework to automate patch diffing and N-day PoC generation. • Donghyeon Kim & Prof. Insu Yun (KAIST Hacking Lab) break down an autonomous agent that tackled CODEGATE 2026 challenges. 🚗 Automotive, Hardware, & Infrastructure (OT) Security • Dr. Camille Gay & Alina present exclusive car hacking and vehicle network research. • Edward Lee (OPCIA) demonstrates how to extract operational tags from proprietary OT traffic and detect unauthorized PLC modifications. • Iceman dives into signaling vulnerabilities, RFID exploitation, and new analysis hardware. Day 2: Hands-On Interactive Workshops • Bootrom Dumping: xenocidewiki (Secret Club) leads a hardware workshop on dumping encrypted pet firmware from a Tamagotchi Paradise via UART/SWD analysis. • Hardware & RFID: Led by Iceman. • Car Hacking: Led by Dr. Camille Gay & Alina. Event Logistics & Venues • Day 1 (Nov 28, 2026): Main Briefings at Marco Polo Hongkong Hotel (google.com/search?kgmid=%2Fg…) (Centenary Room II-III), including a network buffet lunch. • Day 2 (Nov 29, 2026): Workshops & Village (10:00 - 18:00) at JobsDB by SEEK (google.com/search?kgmid=%2Fg…), PCCW Tower, Quarry Bay. Registration Details Seating is limited. Secure your pass at vxcon.hk. For inquiries, contact [email protected]
1
7
1
10
1,163
微软搞了个专门给 Claude Code、Codex 省钱的开源插件,实测成本降低 18.4%~30.9%。 项目叫 TokenEconomizer,不换模型,也不用修改 Agent 本身。 主要做三件事: 记住经验:同一个项目踩过的坑,下次不再重复 精简工具:自动移除长期不用的工具,减少 Token 消耗 打断死循环:发现 Agent 反复报错,提醒它换个思路 支持 Claude Code、Codex、Copilot CLI 和 Pi。 微软在 277 个 SWE-Bench 任务上测试,5 组模型配置全部降低了平均成本,其中 4 组任务成功率还有提升。 我觉得最实用的是打断死循环。AI 有时候一个 Bug 能反复折腾几十轮,Token 就这么烧没了。 你更希望 Agent 省 Token,还是宁愿多花点钱,确保任务一次做对? github.com/microsoft/TokenEc… #agent #ClaudeCode #Codex #AI #OpenSource
100
258
25
2,355
221,225
It's been 7 years from GPT-2 to GPT-6, and a lot has changed. The diagram below shows each generation and its key highlights, based on the public information we know. GPT-2 (2019): More like a prototype to validate the potential of the decoder-only transformer. This was a 1.5B-parameter model that only went through pre-training with next-token prediction. The trained model was only capable of continuing text, not answering questions. GPT-3 (2020): The same recipe as GPT-2, just much bigger and trained on more data to test how far scale alone could go. GPT-3 had 175B parameters, more than 100 times the size of GPT-2, and still only went through pre-training. The scale allowed it to show in-context learning capabilities. When shown some examples, the model could follow them. GPT-4 (2023): On top of pre-training, OpenAI fine-tuned it with RLHF (reinforcement learning from human feedback), so it was not just a completion model but could actually answer questions. It was also multimodal, meaning that it could process input images as well as text, and it handled up to 32K tokens of context. GPT-5 (2025): The first GPT generation with built-in reasoning. Given a prompt, tools, and an effort level, the trained model could generate reasoning tokens before the answer. Its context window grew to 400K tokens, with a router in ChatGPT to decide when a prompt needed thinking mode. GPT-6 (2026): This new model has a bigger context window (1.05M), supports async tool calling, and is generally safer. An interesting feature, called mid-turn steering, is that it can take a new instruction while it's thinking.
29
94
9
874
98,185
中国社区逆向拆解了Claude的封号机制。以下是他们的发现 Claude使用积分制。注册时账号获得初始分数,之后根据行为加分或扣分。低于阈值直接封号。回分很难 影响分数的因素: 注册 > 干净IP是最重要的因素。不一定要家宽,但必须干净 > Gmail已经不是最优选,有权重更高的邮箱服务 > 注册时直接弹KYC的号可以扔了 > 用手机App注册会加分 支付 > 刚注册就订阅会扣分 > 直接上Max会大幅扣分。最好等两天先订Pro > 支付时弹KYC说明账号在考察期,但不一定会封 > 国外实体银行卡加分。Google Play也行。虚拟卡扣分 使用 > 不要一次把限额用完。宁可两个号轮换也不要一个号撸到底 > Desktop客户端有少量加分。尽量少用网页端 > 手机端也有少量加分 > 开启记忆功能加分。记忆文件越多分越高 IP和设备 > 稳定性是关键。不要跳IP,尽量一直用同一个 > 代理必须全局模式并开启TUN > 多设备登录大幅扣分。理想状态是一台电脑一个客户端 > 频繁换IP、频繁登录、高并发都会大幅扣分 账号年龄 > 老号有巨大的容错空间。2023年的号明显更高 > 分数随时间增长,但很慢 其他 > Windows和macOS没区别。Linux会扣分 > 时区和语言基本不影响。不会因为北京时间就封号 > 跟Claude说自己是中国人没问题,记忆不会触发封号 > 大量使用Fable会扣分 > 升级Max大概率进入观察期,除非是高权重老号 观察期和封号 > 接近阈值时账号可能进入观察期。典型特征是掉登录 > 观察期继续扣分就会封号 > 高分号可以扛过多次观察期 关注我,拆解你每天在用的服务背后到底怎么运作的
这篇估计没多少人会认真看完,我就当给未来的自己留个记录。几年后再回来看看,今天我们是怎么给 AI 付费、怎么搭自己的 AI 工具栈的,可能会比现在更有意思。
537
1,353
314
14,439
2,119,515
终于找到克制Codex过度防御的办法了。 用Codex最烦的就是改个小功能,塞满判空等一堆测试,浪费时间。 开源项目OpenCode 在 AGENTS.md 里立的 4 条规矩,专治 AI 的过度设计,建议直接抄进你的AGENT.MD: 1、防过度防御:几乎不会发生的边界/并发问题,讲清概率并经用户确认后再加,严禁为“理论上更稳”私自膨胀代码。 2、入口单次校验:外部数据只在边界校验一次,内部靠类型系统流转,禁止层层重复判断。 3、拒绝提前抽象:单次使用的逻辑就近内联,不随意提前拆分无意义的 helper。 4、控制复杂度:非必要不用 try/catch,能单函数清晰搞定的别硬拆。 以上4条分别对应图片中标记出来的几行内容,直接复制英文即可,github项目地址我帖在评论区。
69
151
12
1,424
157,017
Yup, reverse the chain is fun 👍🏻🤟🏻 @orange_8361
RCE on Windows Calculator is so funny too Nice bugs
2
16
1,312
adr retweeted
Landing within the next 2 hours
499
120
169
3,809
1,617,361
Unprompted , it is really enjoyable
4
1
18
1,392
Astra (launch day) vs Astra (Today). Same prompt. Today's result looks off. The photo realism is gone. It clearly looks nerfed. GPT-5.6 Sol is unusable for me right now too.
421
322
176
6,528
1,190,680
adr retweeted
Replying to @thsottiaux
Look! Tibo
3
2
110
5,860
This is China
9月5日,China GT上海站赛车因撞击起火事件视频流出: 视频中清晰可见撞车起火以及对手车手下车紧急救援事故车辆的全过程。
1
2
521
Everyone is running someone else's harness. Kerness is the kernel underneath harness: provider transport, tool dialects, the turn loop, access control, memory, compaction, resume. You write the Markdown saying how yours behaves. 60 seconds, not 6 weeks. Written in rust with python binding. github.com/xwings/kerness
13
38
2,494
We're excited to highlight 2026 Microsoft Most Valuable Researcher Angelboy (@scwuaptx), who presented at Black Hat USA this week. Drawing on years of Windows security research, Angelboy shared how vulnerabilities can emerge from unexpected interactions between drivers, transports, and system components, even within attack surfaces that have been studied extensively. His talk discussed a new way of thinking about vulnerability discovery, demonstrating how researchers can uncover overlooked paths to privilege escalation and sandbox escape by examining systems across layers rather than in isolation.
1
15
87
8,042
So glad to meet Taiwanese old friend @ashl3y_shen at Google Cloud event during Black Hat USA, with a new collection to have myself version card, so nice! 🥳 – at Las Vegas, NV
2
1
1
19
3,577
My Windows reverse engineering and exploit research workflow has been: 1. Pick a binary to research like tcpip.sys 2. Use github.com/joshterrill/post-… to automate seeing existing binary versions, download, and generate diffs from them 3. Load the resulting .binexport's and .bindiff into an LLM and ask it to analyze 4. Look up the build number of previous Windows version that old binary existed in from uupdump.net/ such as 26100.8328 and create a VM from it 5. Write code and test, working backwards from LLM analysis
14
185
4
1,272
75,303
New #TinyTracer (4.0) is ready: github.com/hasherezade/tiny_… - refactored for compatibility with the latest PIN - and with some new features!
2
28
1
95
9,494