⚡Megatron LLM Hub
全部
Twitter

推特安全流

头条7

GPU进程借WindowServer逃出沙箱执行操作,PoC已公开,旧版本仍可利用。

Chrome's GPU sandbox could TYPE on your Mac. @pwn_ai found CVE-2026-106378: GPU-process code used WindowServer to inject keystrokes, open Spotlight, and launch apps outside the sandbox. PoC: Calculator. Patched in Chrome 155. https://t.co/kX7oMsMGYB #Chrome #macOS #ExploitDev https://t.co/ZazZpYc4fp

浏览器沙箱逃逸cvepoc @cr3ghost 原文 ↗

通行密钥并非不可窃取,攻击链与工具已公开,无密码方案需重新评估。

Think passkeys are the ultimate security solution? Think again. A simple userland malware can compromise your Google Password Manager and potentially all your accounts. Discover the full research here: https://t.co/uuKONgbDCK By 0rcruxe

凭据窃取passkeypoc恶意软件 @28zaaky 原文 ↗

AI 编码 IDE 插件供应链被污染,哈希已公开可自查

RT @marius_benthin: @cursor_ai updated the IOCs on Oct 30, 2026, at 11:30 a.m. PT. They corrected one incorrect hash (as mentioned before) and now state that both versions 2.1.312 and 2.1.313, distributed through Cursor's extension marketplace, contain the malicious payload. https://t.co/6OMFSgwgAV

供应链ai_agent事件poc @cyb3rops 原文 ↗

CVSS 9.8且利用已成型,Ubuntu与Red Hat尚未发布修复内核。

RT @DarkWebInformer: 🚨 Critical Linux kernel TCP flaw enables remote root code execution; CVE-2026-98432 A race condition in the Linux kernel’s IPv4 TCP implementation can trigger a use-after-free before application authentication, potentially allowing an unauthenticated remote attacker to leak kernel memory, corrupt memory or crash the system. Anthropic says Claude developed a working exploit that executed attacker-chosen commands as root against Linux servers with a Multipath TCP listener. The exploit took approximately 53 minutes to complete during testing and has not been publicly released. ⠀ • CVSS: 9.8 Critical, according to the Linux kernel CVE team • Affected: Linux kernels beginning with version 4.4 • Fixed: 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3 • Attack requirements: Network access, no credentials and no user interaction • Exploitation status: No known attacks in the wild ⠀ Ubuntu and Red Hat Enterprise Linux had not released corrected kernel packages as of October 29. Debian, Oracle Linux, Amazon Linux and SUSE have released fixes for at least some affected branches. Distribution severity ratings differ substantially. Red Hat rates the flaw 5.9 because it considers the race difficult to win and describes denial of service as the most likely practical impact.

cverce内核poc @0xocdsec 原文 ↗

给出可照做的内核版本核对与升级路径,覆盖主流发行版。

RT @derekmross: 🚨 PSA for anyone running a Linux server, node, or relay CVE-2026-98432 is a critical Linux kernel bug (CVSS 9.8). Someone on the internet can send crafted TCP packets and take over your machine. No login needed. Ubuntu and Red Hat haven't shipped a fixed kernel yet. Debian, SUSE, Oracle and Amazon have. What to do today: 1. Check your kernel: uname -r Fixed versions: 6.1.190+, 6.6.159+, 6.12.113+, 6.18.56+, 7.2.10+ or 7.3+. Below that, you're exposed. 2. If your distro has a fix, update and reboot. 3. If not, block the attack. It needs TCP packets with "IP options," which normal traffic never uses. One nftables rule drops them. Log what gets dropped so you can see if someone's probing you. 4. Turn off MPTCP unless you need it: sysctl -w net.mptcp.enabled=0 Programs built with Go 1.24+ use MPTCP automatically, and that's how the known exploit gets in. 5. Make it survive a reboot, and check it still works after one. If you run a Bitcoin node, a Lightning node or a Nostr relay, your machine is on the internet by design. Do this today. Self-hosting is freedom. Patching is part of the job. 🛡️

cve内核rce @0xocdsec 原文 ↗
必看19

有效 EV 签名可绕过信誉拦截,该证书指纹应立即进入检测黑名单。

EV "Hangzhou Yueque Li Technology Co., Ltd." signing two Gh0st RAT variants. h/t @malwrhunterteam First: Golden Gh0st RAT from GoldenEyeDog: https://t.co/90Wd9NzpFd (Image: Golden Gh0st Rat lure from July) Second: GodRat: https://t.co/UvQYYLM1gY https://t.co/B01Ko6Iox0

恶意软件代码签名事件 @SquiblydooBlog 原文 ↗

LLM 驱动的逆向已达字节级还原,自建 Agent 可直接借力二进制分析。

RT @r00tth3w0r1d: Trick or treat? We brought source code. 🎃 Using REA (@reatools), we reconstructed its C/C++ source from the original EXE — 270/270 authored functions, 84,968/84,968 bytes, 100% byte-exact. Code: https://t.co/N6npFsx7XB Video: https://t.co/naMjkgoJxn What's next? Let's REA it! https://t.co/1x4oWsK9Jo

ai_agentllm逆向工具 @morluto 原文 ↗

任务下发与结果回传可走不同通道,流量侧检测特征随之整体改变。

Halloween Update 🎃👻 Added numerous new features related to operational security, introduced WebSockets as a communication channel, added file based P2P capabilities, improved the user interface, and fixed bugs reported by customers found within the malleable C2 profile system. Another exciting addition is independent ingress/egress listener configuration, allowing operators to use separate communication channels for task requests and returning execution results simultaneously! I recorded a small video demonstration of said feature. YouTube Link below🔗

c2工具对抗 @C5pider 原文 ↗

若属实则边界隔离假设失效;目前只放出了代号与哈希,需要跟进细节。

RT @ikotas00: We discovered a critical vulnerability that allows remote attackers to bypass firewall restrictions and establish bidirectional communication. Codename: BunkerBuster💥 eeabb157b3a3f251b52656c1a445564c8f7a4cb2340f0d41de76d2cbc542cde6 https://t.co/Wh07O75MLa

绕过防火墙事件 @0xocdsec 原文 ↗

多家 AI 开发平台被入侵却无 IOC 公开,同类自建环境无法自查是否中招。

I think it is incredibly important to publicly share IOCs from these compromises(railway, opencode, fireworks, northflank, etc), or atleast share them privately among major companies. this is important information to understand attackers MO to defend against similar attacks. unfortunately, this seems not happening.

供应链事件ai_agent @S1r1u5_ 原文 ↗

AI 编码扩展经非官方市场分发被投毒,插件供应链防护需覆盖这些渠道。

RT @MsftSecIntel: On October 29, 2026, Cursor reported that a tampered copy of Claude Code extension (anthropic.claude-code, version 2.1.313 and 2.1.312) was available and served on the Cursor extension marketplace for a few hours. https://t.co/AOSQOdtjFU Microsoft has confirmed that the malicious packages found on Cursor marketplace were never published nor distributed from the Microsoft Visual Studio Marketplace extension and the npm anthropic-ai/claude-code, which were not affected by this incident. Anthropic confirmed to Microsoft that their systems were not compromised and they did not develop the malicious package posted in Cursor marketplace. At this time, all evidence suggests that the incident is limited to the Cursor marketplace. Microsoft Threat Intelligence analyzed the JavaScript embedded into the malicious extension package. The injected loader runs when the extension module loads and attempts to add an attacker-controlled Secure Shell (SSH) key, enable the local SSH service, establish reverse SSH forwarding, configure persistence, and report host details. Successful remote access depends on host privileges, SSH configuration, and the attacker’s server configuration; however, the loader may still write keys and persistence artifacts even when tunnel access fails. Read the blog for technical analysis and guidance on identifying affected systems, investigating endpoint and network activity, and removing confirmed malicious SSH keys and persistence artifacts.

供应链ai_agentllm @artem_i_baranov 原文 ↗

把AI能力直接嵌进C2工作流,示范了模型如何降低后渗透操作门槛。

RT @_CobaltStrike: New Blog: Based on his talk at Black Hat, @0xTriboulet discusses integrating Windows AI/ML APIs into Cobalt Strike’s workflows and presents proof-of-concept implementations for AI-augmented post-exploitation capabilities in Cobalt Strike. https://t.co/B8icpllZHV https://t.co/JIcw0JIHpe

ai_agent后渗透c2poc @G3tSyst3m 原文 ↗

覆盖全部 Go 1.24+ 服务端,竞态难赢,可据此判断加固优先级

regarding cve-2026-98432: it is a uaf in tcp, mptcp is needed to make the race win and to get the leak and write. the impact of the bug is bad, its essentially go2shell, caddy2shell etc. go 1.24+ servers are all vulnerable, because mptcp is enabled by default. BUT: 1. It is extremely hard to win the race reliably, anthropic said it took 40 mins for them to get leak and defeat kaslr 2. It is probably hard to exploit in practice. most routers, lbs and isps drop ip options, without that header you can't race at all, so realistically only a same network attacker can try, which is pretty common in sandbox infra my mythos worked on this for last 3 days burned ~$2k and couldn't build a reliable poc that can be sprayable. gpt-5.6 sol als exhausted my subscription and couldn't even get a leak. either anthropic has a unreleased model which is better than their public one or they have a pro exploit dev internally.

rce内核cveai_agent @S1r1u5_ 原文 ↗

网络可达即可利用,自建分析库常暴露在内部网。

RT @cvedotnow: 🔴 ClickHouse: remote code execution in database server (CVE-2026-116366) An unauthenticated attacker on the network could run arbitrary code on the server under certain conditions. Patch → 26.3.62.3-lts, 26.8.36.3-lts and more https://t.co/lRVKJD1G77

cverce数据库 @0xocdsec 原文 ↗

第三方复核指出堆偏移令利用依赖构建,版本号判断并不可靠。

RT @devploit: Thanks @TenableSecurity for the mention! They used my php2shell PoC to reproduce CVE-2026-109618/109619 on several PHP 8.4/8.5 setups and wrote up why the exploit is so build-dependent: heap offsets, worker reuse, and why version checks lie. Great read. https://t.co/8TY4OAPV6x

cvepocrce @0xocdsec 原文 ↗

官方内核补丁未出前可直接粘贴执行的过渡缓解方案。

RT @r0ckstardev: 🔴 If you have a VM running Ubuntu and want to patch the CVE-2026-98432 Critical Linux Kernel Flaw, here is the script I used to secure my servers: # CVE-2026-98432 stopgap. For Ubuntu 20.04 and 24.04. # Log in as ubuntu, copy everything in this file, paste it into the SSH session. # Done when the last line prints: active sudo apt-get update -qq </dev/null && sudo apt-get install -y nftables </dev/null sudo tee /etc/nftables-cve-2026-98432.nft >/dev/null <<'EOF' add table inet cve_2026_98432 flush table inet cve_2026_98432 table inet cve_2026_98432 { chain pre { type filter hook prerouting priority raw - 10; policy accept; ip hdrlength > 5 meta l4proto tcp counter drop } } EOF sudo tee /etc/systemd/system/cve-2026-98432-nft.service >/dev/null <<'EOF' [Unit] Description=CVE-2026-98432 stopgap: drop IPv4 TCP with IP options After=https://t.co/yv6n00FLOm Before=https://t.co/w7O0ZyUiWr docker.service [Service] Type=oneshot RemainAfterExit=yes ExecStart=/usr/sbin/nft -f /etc/nftables-cve-2026-98432.nft [Install] WantedBy=https://t.co/XN4fnfddjw EOF sudo systemctl daemon-reload && sudo systemctl enable --now cve-2026-98432-nft.service && systemctl is-active cve-2026-98432-nft.service

cve内核缓解 @0xocdsec 原文 ↗

Ubuntu 修复仍在测试未发布,细节已扩散,暴露 MPTCP 的主机无补丁可用。

RT @r0ckstardev: I spent the last few hours going down the CVE-2026-98432 Critical Linux Kernel Flaw rabbit hole 🕳️ and here are a few thoughts: - It is legit and scary. Kinda crazy that it's been known for 2 weeks & there is still no patch for popular distros (the Ubuntu fix, for example, is developed but not publicly released, still in testing). - MPTCP needs to be enabled, and on some unorthodox port. It seems that ISPs globally have started blocking options on 22, 80, 443 and other popular ports (good!). - If you're running a server with valuable data, disable MPTCP and/or IP options NOW. They are not necessary in 99% of use cases anyways. - It takes about 50 minutes on average to gain access to a server if all conditions are met. My servers started getting hit today; luckily the script I shared earlier handled the situation and rejected the probes. Stay safe & enjoy the wild AI psychosis ride we're in.

内核cvemptcp @0xocdsec 原文 ↗

共享链接是低门槛投递面,会话内执行意味着私聊与账号一并失守,手法可迁移到同类 AI 前端。

RT @ehsayaan: We hacked ChatGPT at @Parameter_ai. Opening a malicious shared ChatGPT link could leak your private chats and compromise your ChatGPT account. The link looked like a shared conversation. Opening it could run attacker-controlled code in your signed-in session, giving it access to chats you’d never shared. OpenAI rated the issue as P1 and promptly patched it. Full write-up: https://t.co/pF7VoCjKS5

ai_agentllmxss @0xocdsec 原文 ↗

邮件确实来自 Google 域名并通过认证,常规发件人校验完全失效,仿冒极难识别。

RT @TrevorCampbell_: Just got a phishing email that Gmail authenticated as Google. DKIM/DMARC passed. I can see a lot of people falling for this one, since it plays to the idea that your account has already been compromised. Scammer puts the scam in an app-password name, triggers a real Google alert, then forwards the alert email to your email address. How this attack works, and why it’s convincing: 1. The trick: The attacker puts the threatening text and shortened URL into an app password’s customizable name on another account. Google incorporates that text into a real security notification and signs the email. 2. The delivery: That notification is forwarded through another mailbox to the victim. In this message, Gmail still recorded passing Google DKIM and DMARC authentication. 3. What makes it convincing: The Google sender, signature, branding, and “Check activity” button are legitimate. The suspected attacker-supplied text appears inside the same trusted-looking template, using a deadline to push you toward its separate link. 4. The giveaway: Both the notification and its genuine Google button refer to someone else’s account. The “permanent access after 24 hours” claim is false; app passwords can be revoked. Be careful out there folks!

钓鱼绕过邮件 @0xocdsec 原文 ↗

AI 与经典 fuzzer 形成互补而非彼此替代,可直接用于调整漏洞挖掘的投入分配。

RT @cr3ghost: Google Project Zero pitted AI against a V8 fuzzer. AI won. Barely. 6 months: BigSleep (AI): 45 actionable vulns Fuzzilli: 37 Only ~25% overlap. The surprise? Each found bugs the other missed. Research by @5aelo: https://t.co/Akcg84p4UF #Fuzzing #AI #V8 https://t.co/n9mF2j7uhl

fuzzingllm浏览器报告 @cr3ghost 原文 ↗

用云端 AI 做漏洞研究可能使工作成果被纳入模型训练,敏感材料应留在本地模型。

RT @seanhn: haha, so apparently OpenAI are about to start training on your Daybreak Blue sessions, with no way to opt out. Literally stealing people's private work, as they do it, is cartoonishly evil. It also will presumably kill Daybreak usage by security researchers. Or at least, it should. The thing OAI have had to swear they are not doing to math researchers, they are now explicitly about to start doing to security researchers. Yet another reminder of the importance of open models.

llmai_agent数据隐私 @mrexodia 原文 ↗

形式化验证从内核延伸到用户态真实二进制,对漏洞根除思路有长线影响。

RT @v12sec: x86 formal verification is coming! The (first?) published machine-checked proof of a security property for the end-to-end execution of a Linux process running an x86 binary statically linked against glibc. And we believe it scales! A joint research project with @zellic_io 🧵👇 https://t.co/719dSrp0BG

形式化验证二进制报告 @alkalinesec 原文 ↗
推荐15

无需注入器即可落地载荷,是低成本持久化与规避检测的成熟手法。

🧠 What if you never needed a DLL injector? DLL proxying turns a target's own dependencies into the execution vector, slipping past conventional detection routines. 👉 https://t.co/WYKaC84lhS https://t.co/S8g34zZdj6

绕过windows工具 @GuidedHacking 原文 ↗

现成检测查询与规则可直接落地到Defender XDR环境。

🎣 𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗧𝗲𝗮𝗺𝘀 𝗩𝗶𝘀𝗵𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗗𝗲𝗳𝗲𝗻𝗱𝗲𝗿𝗫𝗗𝗥 ​ MC1532241 caught my attention, so I started exploring how the new 𝙲𝚊𝚕𝚕𝙰𝚌𝚝𝚒𝚟𝚒𝚝𝚢𝙴𝚟𝚎𝚗𝚝𝚜 data could help defenders investigate suspicious Microsoft Teams calls. https://t.co/jZczlsZkKU​ I’ve put together a KQL hunting approach that correlates: ​ 🔹 𝚂𝚎𝚌𝚞𝚛𝚒𝚝𝚢𝚁𝚒𝚜𝚔𝙸𝚗𝙲𝚊𝚕𝚕𝙳𝚎𝚝𝚎𝚌𝚝𝚎𝚍 signals from 𝙲𝚕𝚘𝚞𝚍𝙰𝚙𝚙𝙴𝚟𝚎𝚗𝚝𝚜 🔹 The associated Teams call/meeting activity in 𝙲𝚊𝚕𝚕𝙰𝚌𝚝𝚒𝚟𝚒𝚝𝚢𝙴𝚟𝚎𝚗𝚝𝚜 🔹 𝙲𝚊𝚕𝚕𝙸𝚍 as the correlation point between the two ​ The goal is to give defenders additional context when investigating potential 𝘃𝗶𝘀𝗵𝗶𝗻𝗴 𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲-𝗯𝗮𝘀𝗲𝗱 𝘀𝗼𝗰𝗶𝗮𝗹 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗮𝗰𝘁𝗶𝘃𝗶𝘁𝘆 𝗶𝗻 𝗧𝗲𝗮𝗺𝘀. KQL Code: https://t.co/aJuJ4GAcGV ​ #DefenderXDR #MicrosoftTeams #KQL #ThreatHunting #Vishing

检测kql钓鱼 @0x534c 原文 ↗

企业终端可被该类程序绕过检测,需核对 Defender 平台版本

Microsoft has fixed a five-year-old bug in Defender. This bug allows programs classified as PUAEnterprise to run on some enterprise devices without being blocked or detections Up to &gt;= 4.18.26100.4 (Sometimes contacting support is more efftective than submitting an issue to MSRC) https://t.co/GnZ62QYaDC

绕过移动端cve @anylink20240604 原文 ↗

自托管红队实验环境,含 Apache 重定向器与 Windows/Kali 主机

redStack — Terraform-deployed red team operator lab on AWS: Mythic, Sliver, and Adaptix C2 behind a production-style Apache redirector, with Kali/Windows hosts and a Guacamole portal. Built for authorized training and self-hosted cyber ranges. https://t.co/IqIfjiE3zj https://t.co/uyPVZglW75

工具云提权 @KitPloit 原文 ↗

边界设备集中被利用,其攻击面选择对自查有参考价值

RT @UK_Daniel_Card: This is super useful (and cool from a CTI pov) (obviously incidents for orgs/ people are not fun) I wrote a blog last week saying mass exploitation of everything wasn’t likely but we would probably see limited exploitation of appliances! Turns out my cyber tea leaf 🍁 reading wasn’t awful! Great job from the @Huntio team! #PHP2Shell

供应链事件报告 @cyb3rops 原文 ↗

开源 Android RAT 可直接上手,适合授权测试或了解移动端控制面

Lab-RATS-v1.6 — Lab-RATS is a powerful and lightweight Android Remote Administration Tool full of features that enables remote monitoring, management and interaction through a sleek web interface and supports the newest modern Android releases. https://t.co/tlLQn3iVnB https://t.co/gevvwdFsm9

工具移动端cve @KitPloit 原文 ↗

该发行版需先测内核或启用缓解,官方称受影响范围明确。

RT @AlmaLinux: ⚠️ Linux kernel vulnerability CVE-2026-98432 is a use-after-free in the code that sets up incoming IPv4 TCP connections. AlmaLinux 8, 9, and 10 are affected. Patched kernels are available for testing! https://t.co/tywvaD8LDZ

cve内核 @0xocdsec 原文 ↗
更多49

号码租用产业链被端,印证短信验证码仍是账号接管的主要路径。

RT @DarkWebInformer: 🚨🇪🇺 In October 2025, Europol helped dismantle a massive cybercrime-as-a-service network that had enabled fraud and other criminal activities across Europe. The operation, known as SIMCARTEL, resulted in 7 arrests and the seizure of 1,200 SIM box devices, 40,000 active SIM cards, and 5 servers. The network had provided criminals with phone numbers from more than 80 countries, enabling the creation of over 49 million online accounts used for phishing, impersonation, and financial fraud. 🎥 This is footage from the police raids carried out during the operation.

事件犯罪基础设施账号接管 @NetAskari 原文 ↗

影响Claude生成内容的可追溯性,依赖Claude产出的流程需评估水印残留。

RT @petesena: Claude gonna start snitching on Tuesday. Anthropic just emailed customers. Starting Nov 3, text carries an invisible watermark. You can't see it. It survives copy and paste, and it can survive some editing. Only regulators, journalists, fact-checkers and researchers can check for it right now. It doesn't say who you are. It says Claude touched the words. Even a quick proofread counts. If your brand voice runs on Claude, write like someone might check. Because now someone can.

llm水印溯源 @artem_i_baranov 原文 ↗

描述了自动化混淆资金路径的手法,对链上追踪规则设计有参考价值。

RT @chainalysis: Our investigators are observing never-seen-before laundering patterns with the stolen funds from the Binance exploit. Thousands of transactions in a short period suggest that the attackers are using automation to obscure the trail of funds. Rather than simply moving funds onward, they repeatedly circulate them among addresses, often sending funds back to addresses involved earlier in the sequence. Read more: https://t.co/PJ71R2dxnS

事件洗钱区块链自动化 @artem_i_baranov 原文 ↗

轻量PowerShell脚本可减少Windows遥测与预装组件带来的暴露面。

Win11Debloat v2026.10.30 — A simple, lightweight PowerShell script that allows you to remove pre-installed apps, disable telemetry, as well as perform various other changes to declutter and customize your Windows experience... https://t.co/e3czWDnmi5 https://t.co/8KmDeO2fCA

工具windows遥测 @KitPloit 原文 ↗

RT @Hannibal9972485: Binance 🚨🚨is balance sheet restructuring & trying to stop a bank run that was likely already forming due to their internal bad debts or fractional reserve practices. . Binance is facing a massive liquidity crisis (maybe due to bad loans, FTX-style internal trading desk losses, or regulatory freezes). They need to lock up $6.3 billion in user funds without admitting insolvency. 🚨 They engineered a "hack" using the EIP-7702 vector. They move the funds to wallets they control. They freeze $3B of it (to stop the bank run on their own exchange) and let $3.3B "escape" into dormant BTC wallets and Uniswap swaps (which are likely wash trades back to their own market makers). They crash the market 4-5% to liquidate short positions and scoop up cheap assets. 🚨 Look at the forensic breakdown of the attack. In the span of 20 minutes, the attacker drained funds across ten different blockchains, including Bitcoin, Ethereum, Tron, Solana, XRP, Optimism, and Avalanche. 🚨 These blockchains do not share the same cryptographic curves or key types. Bitcoin uses ECDSA (secp256k1). Solana uses Ed25519. XRP uses its own variant of ECDSA. A single compromised laptop does not possess the magical ability to sign valid transactions across ten fundamentally different cryptographic architectures in 20 minutes. As TRM Labs explicitly noted in the forensic data, a single phished Ethereum authorization cannot explain Bitcoin, Tron, XRP Ledger, Solana, and layer-2 transfers all occurring simultaneously 🚨 If a real hacker compromised their laptop, they would have to navigate Binance’s internal infrastructure to access 10 separate, isolated hot-wallet signing services, bypassing hardware security modules and multisig requirements for each individual chain. That takes hours or days of lateral movement, not 20 minutes. 🚨 The only entity capable of executing simultaneous, validly signed withdrawals across ten different chains in 20 minutes is the exchange’s own internal automated withdrawal engine —🚨 According to the on-chain data, the attacker took 325 million USDT and swapped it for Ethereum on UniswapX at a 10 percent discount, paying a massive premium for speed. Professional market makers absorbed that trade. Why would an external hacker take a 10 % haircut on an irreversible, unsanctioned swap when they control the wallet? They wouldn't.⚠️⚠️🚨 Furthermore, the Bitcoin stolen was split into exactly 1,000 BTC chunks across 17 new wallets and then sat completely dormant. 🚨 It was not laundered. It was not moved to a mixer. It was organized into clean, institutional-sized blocks and parked. Hackers do not park stolen funds in neat, 1,000 BTC increments. Institutional custodians do this when they are restructuring balance sheets or moving funds to cold storage under their own control.🚨 The attacker even routed test transactions back to itself before the main sweep. This is the behavior of an internal engineer testing a deployment script.🚨 —- Binance is doing exactly what FTX did before its collapse, just with a more sophisticated smokescreen. They have halted withdrawals. They are promising to make users whole with money they do not have. 🚨 Within the next 30 to 60 days, Binance will slowly resume withdrawals for small retail users to maintain the illusion of solvency. Large institutional whales will find their accounts frozen indefinitely pending "investigation" into the hacked funds. The $3.3 billion in Bitcoin and Ethereum currently sitting in dormant wallets will not be recovered. It will be slowly commingled with other exchange flows over the next year to quietly plug Binance's internal holes. 🚨 You are watching a bank robbery designed to hide the fact that the largest exchange on earth is out of money. - https://t.co/rXv2cIejt8 https://t.co/bnfBLIMvFo

事件绕过 @cyb3rops 原文 ↗

牵涉Agent调用与对话数据在云端被留存的边界。

RT @lukOlejnik: OpenAI reversed the plan to force training AI models based on usage without "off" option". Always amazing to see technology companies rediscovering what other technology companies discovered the hard way in previous 20 years ;) Speaking from PR/communication experience here. https://t.co/FfYaaKYHlO

llm隐私 @0xocdsec 原文 ↗

Agent自主发起循环调用会造成无上限的成本消耗。

RT @krunkosaurus: DeepSeek-V4.1-Flash-1017 is good but I suddenly agree its clearly no where close to frontier. I was working on a complex Sogni video project overnight and to check remaining balance it set 1-min cron to generate a 4 second clip—I dunno why it thought that was the best way to check balances... I didn't realize it until the next morning where it blew through its entire $170 left in that specific user account. Super disappointing. Any Anthropic or OpenAI model would have been far more careful, and I'm not even running abliterated models. I also feel GLM5.3 Flash would not have done such thing. It just seems more of a careful model in my testing.

ai_agentllm事件 @0xocdsec 原文 ↗

RT @jino_rohit: im building noro - a image and video inference library. it is built on top of huggingface diffusers with a focus on performance. i currently have the krea-2 model implemented and across all image dimensions, on average we have a 20% speedup. im currently building it for nvidia chips and also extending to amd as well. all the kernels and tricks are written in triton and gluon! repo link in the comments.

工具推理 @morluto 原文 ↗

有国家背景的雇佣间谍软件仍在活跃投递,说明零点击攻击链尚未被封堵。

RT @IntCyberDigest: 🚨 BREAKING: Apple has warned iPhone users in 99 countries that mercenary spyware is targeting them. Lebanese journalist Radwan Mortada is one. Apple had already warned him in 2021. He says he doesn't know if this attack is linked to his film on Lebanon's 2024 pager explosions. https://t.co/oqVBxYHotJ

移动端间谍软件事件 @0xocdsec 原文 ↗

VTL1 隔离是 Windows 提权与凭据窃取的主要障碍,理解边界才能找绕过点。

Isolated User Mode (IUM) is the VTL1 usermode. Processes running there are called trustlets (LsaIso.exe, BioIso.exe) IUM prevents VTL0 from reading or writing memory, even from kernel drivers. It also restricts thread access, debugging, handle creation, DLL loading, and more. https://t.co/oGEZs8cAja

内核绕过windows @papadelospapas_ 原文 ↗

供应商自身的运维失误会直接导致邮件与云账号被劫持,是常被忽视的攻击面。

RT @foilmanhacks: The CEO of Israeli spyware company 'Discover-Tech' has let the domain for their primary email address lapse. Discover-Tech is the same company behind the 'Adfection' campaign. This could allow attackers to hijack their email address, Microsoft account and Dropbox. https://t.co/A3KmIi6K3c

事件供应链osint @0xocdsec 原文 ↗

RT @RichardSSutton: One reason I am sanguine about the technological singularity is that I have seen it coming for over 50 years. In 1975, when I was 17, I wrote: "Before man can most effectively aid his problem solving machinery, he must first understand it. At present, his understanding is very poor. However, when the breakthrough in understanding the brain occurs, improvement will come very fast and in an exponential increase. Once man has created an intelligence greater than his own, that intelligence should be able to create an intelligence greater than itself, as man whose intelligence is less than its was able to create its intelligence. Thus, the level of intelligence will shoot up with incredible rapidity, probably within the space of years. Within a few years of a breakthrough, which may occur any time now, man will either have been left behind by a superior intelligence of his own making, which he can no longer understand, or he will have become that intelligence and will no longer be like anything imaginable today."

ai @S1r1u5_ 原文 ↗