⚡Megatron LLM Hub
全部
Twitter

推特安全流

头条8

无需微调即可逆改变本地模型拒答行为。

RT @lordx64: I built a tool to control your models refusals from the KV cache without modifying the weights. Handy for Cyber capable models Phantom-KV: Refusal removal as a loadable KV-cache graft zero weight modification, no refusal-direction projection, fully reversible. https://t.co/KutyA17d2T

llmai_agent工具 @0xocdsec 原文 ↗

含 pre-auth RCE 与沙箱逃逸的完整链条,可直接复现与写检测。

Want to get better at web exploitation? Study @assetnote's research (now @SLCyberSec). • WordPress core pre-auth RCE • ServiceNow sandbox escape • AEM Cloud pre-auth RCE • Discourse cache poisoning + file read • Goja OOB -> RCE Read the chains: https://t.co/INAyI2TUPz https://t.co/drF4Gv3Y1g

rce逃逸web报告 @cr3ghost 原文 ↗

证书与文件名均正常,靠信任判断无法识别,IOC 可直接布防。

Trojanized PeaZip signed "Listenmaa Consulting Oy" DLL swapped in for sideloading, h/t @malwrhunterteam who found it while it was FUD on VT. Certs been reported, getting detection now: C2: apiupdated[.]com/Hrd50dFgjQW53n 157.173.16[.]197 https://t.co/kcyzBwpe3b https://t.co/IomMtKcNEs

供应链dll侧加载恶意样本ioc @SquiblydooBlog 原文 ↗

同一证书已出现第 7 个序列号,签名滥用成规模,IOC 可溯源。

Signed "Lway Firmware" - 7th unique cert serial for them 🫥 h/t @malwrhunterteam Microsoft teams lure -> BabaDeda loader -> CNCMachineRMS RAT C2: wizardexpensive[.]com: 96.62.218[.]59 https://t.co/fs2qVYQHsh https://t.co/Xin8Am7lFO

供应链恶意样本钓鱼ioc @SquiblydooBlog 原文 ↗
必看6

漏洞真实可打,运行 PHP 服务需尽快补丁。

this is a very dangerous narrative. just look at the engagement it's getting. for the first ~10 hours, the challenge setup wasn't even realistic. then the single threaded PHP server got effectively DoSed by traffic, making the challenge practically unsolvable, this is solvable. none of this means the vulnerability isn't exploitable if bad actors have frontier models. patch your vulnerable PHP servers. the bug is exploitable, and it's bad.

rceai_agent事件poc @S1r1u5_ 原文 ↗

新增 PCFG 攻击与 FPGA 上的 bcrypt 破解。

RT @hashcat: hashcat v8.0.0 released! The largest update in the history of hashcat, with PCFG attacks, hash recipes that crack hashes without a hash-mode, and bcrypt on FPGA boards. Release notes: https://t.co/0ffrO586PJ Download: https://t.co/Grn4jZmAWr https://t.co/ir6CSSsGcJ

工具口令破解gpu @Jean_Maes_1994 原文 ↗

完整手法与代码公开,可复现的本地凭据窃取路径。

The project captures the PIN by placing a CPU hardware breakpoint on the function that packages the typed PIN inside LogonUI.exe, reading the argument from the trapped thread context in a Vectored Exception Handler, and letting the original, unmodified instruction execute, so the PIN login still succeeds normally. https://t.co/nyyZQc3zoY

凭据窃取windows工具 @ipurple 原文 ↗

给出在野攻击链、样本哈希与蜜罐关联数据,便于直接排查。

🚨 𝗣𝗛𝗣𝟮𝗦𝗵𝗲𝗹𝗹: 𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝗻𝗴 𝗵𝗼𝗻𝗲𝘆𝗽𝗼𝘁 𝘁𝗲𝗹𝗲𝗺𝗲𝘁𝗿𝘆 𝘄𝗶𝘁𝗵 𝗲𝗻𝗱𝗽𝗼𝗶𝗻𝘁 𝗽𝗿𝗲𝘃𝗮𝗹𝗲𝗻𝗰𝗲 ​ An interesting correlation while looking at the ongoing PHP2Shell activity. ​ The 𝗗𝗲𝗹𝗳𝘁 𝗧𝗵𝗿𝗲𝗮𝘁 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗟𝗮𝗯, running 𝟯𝟬𝟬 𝗵𝗼𝗻𝗲𝘆𝗽𝗼𝘁𝘀, has been capturing real-world PHP2Shell activity, including attempts to move beyond scanning into OS command execution, PHP code delivery, and malware deployment. https://t.co/AzsAQRu8PM ​ One of the associated malware samples has the SHA256: ​ 𝟶𝚌𝟻𝟺𝟾𝟼𝚏𝟶𝟼𝟼𝟸𝟶𝚎𝟸𝟼𝚎𝟹𝟷𝟼𝚏𝟺𝚋𝟹𝟹𝚏𝟹𝟷𝟾𝚋𝟶𝚍𝟼𝟺𝟼𝚍𝟻𝚎𝟷𝟷𝟹𝟹𝟽𝟾𝟿𝟹𝟹𝟷𝚊𝚊𝟸𝚏𝚎𝚍𝟶𝟷𝟸𝟹𝟼𝟶𝚍𝟷𝟺𝟾𝟼 ​ Looking at the same SHA256 in 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗗𝗲𝗳𝗲𝗻𝗱𝗲𝗿, the file prevalence shows 𝟲 𝗱𝗲𝘃𝗶𝗰𝗲𝘀 𝘄𝗼𝗿𝗹𝗱𝘄𝗶𝗱𝗲, with: 🔹 First seen: 𝟮𝟮 𝗢𝗰𝘁 𝟮𝟬𝟮𝟲, 𝟭𝟭:𝟱𝟲 𝗔𝗠 🔹 Last seen: 𝟮𝟯 𝗢𝗰𝘁 𝟮𝟬𝟮𝟲, 𝟴:𝟬𝟭 𝗔𝗠 ​ What makes this interesting is the timing. The Defender observations closely follow the new attack telemetry recorded by the honeypot infrastructure. ​ This doesn't prove that those 6 devices were compromised through PHP2Shell, but it demonstrates the value of 𝗰𝗼𝗿𝗿𝗲𝗹𝗮𝘁𝗶𝗻𝗴 𝗳𝗿𝗲𝘀𝗵 𝗵𝗼𝗻𝗲𝘆𝗽𝗼𝘁 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝘄𝗶𝘁𝗵 𝗲𝗻𝗱𝗽𝗼𝗶𝗻𝘁 𝘁𝗲𝗹𝗲𝗺𝗲𝘁𝗿𝘆 to understand how quickly newly observed threats may surface across environments. ​ 𝗛𝗼𝗻𝗲𝘆𝗽𝗼𝘁 → 𝗜𝗢𝗖 → 𝗘𝗻𝗱𝗽𝗼𝗶𝗻𝘁 𝘁𝗲𝗹𝗲𝗺𝗲𝘁𝗿𝘆 → 𝗧𝗵𝗿𝗲𝗮𝘁 𝗵𝘂𝗻𝘁 ​ That is where threat intelligence becomes actionable. ​ #PHP2Shell #CyberSecurity #ThreatIntelligence #ThreatHunting #MicrosoftDefender #Malware #IncidentResponse #SOC

事件webshellioc蜜罐 @0x534c 原文 ↗
推荐15

独立复核前沿模型漏洞利用能力的公开分析。

RT @LiveOverflow: I took a closer look at the https://t.co/rtT8YaMOFZ CTF, which aims to demonstrate that PHP2SHELL is not realistically exploitable. So I gave it a try. Note: My analysis could be wrong, and I'm more than happy to be proven otherwise. Here’s what I found: • The server responds with Server: nginx and X-Powered-By: PHP/8.4.26. • My differential fingerprinting scanner independently confirmed at least PHP 8.2+ • Hypotheses: I considered two possible setups: ◦ nginx → PHP-FPM ◦ nginx → Apache → PHP-FPM/mod_php • After comparing the server's behavior against local test environments, the closest match appears to be nginx → PHP cli-server. Plot twist: the response behavior is consistent with the PHP cli-server serving a static index.html rather than an index.php. If that's indeed the case, the challenge appears to be intentionally made unexploitable. My conclusion (for now) is that I'm questioning whether this is a good-faith challenge. Happy to be corrected if I've missed something!

rcephppoc报告 @artem_i_baranov 原文 ↗

可脚本化分析内核崩溃转储的新工具。

RT @DumpAnalysis: The initial version of the Prolog WinDbg extension is out there: https://t.co/4xMLzwpT1y The accompanying book will be published later today. Finally, I resumed work that I started more than 15 years ago... https://t.co/EYJktc2hxN

工具内核windows @kernullist 原文 ↗

本地服务可被直接转成 Agent 训练与测试环境,权限面外扩。

RT @earakely: Its never been easier to make rl environments: "Make an environment from my local Grafana at http://127.0.0.1:3000. Cover dashboards and folders, including creating, renaming and deleting folders." -> https://t.co/vv97cZlPgm

ai_agent工具自动化 @morluto 原文 ↗

兼容哥斯拉、冰蝎等现成管理器,内存马投递门槛进一步降低。

NMG generates parameterized .NET memory shell payloads that inject into IIS/ASP.NET applications without leaving files on disk. After injection, you connect with existing webshell managers (Godzilla / Behinder / AntSword) https://t.co/iCelqRngY4

工具webshell内存马绕过 @ipurple 原文 ↗

签名者已全部录入 CertGraveyard,可直接转成信任判断与检测规则。

We updated the blogpost. After the original post: - I saw the malware in the network I monitor - Others reached out to me to share TTP - Others have published about it - We've connected it to other signers and malware All the signers have been added to CertGraveyard

代码签名恶意软件检测 @SquiblydooBlog 原文 ↗

主流 Android root 与 hook 方案更新,影响移动端提权与检测对抗。

Magisk v31.1 — Android customization suite providing root access (MagiskSU), systemless module installation, boot image unpacking/repacking, and Zygisk runtime code injection for advanced device modification. https://t.co/cdeBhbfmrI https://t.co/x8eMXuFM7L

移动端root工具 @KitPloit 原文 ↗

现成 BOF 可直插 C2 使用,省去自行实现内存操作的开销。

RT @Dinosn: chryzsh/redteam-bofs: Repository for various Beacon Object Files (BOFs) https://t.co/GoZxs2zcYD

c2工具红队 @0xTriboulet 原文 ↗

安装包解包取载荷的工具更新,便于流水线批量拆样本。

Debloat 1.7.1 - https://t.co/Q10T7Ja25t In 1.7.0, we added a "--no-extract" optional flag. This makes it easier to use Debloat in pipelines that expect only 1 item in return. Without it, it would extract NSIS installers automatically. 1.7.1 fixed the release bins for Tcl9

工具恶意样本 @SquiblydooBlog 原文 ↗

模型侧策略收紧直接压缩自建 agent 能做的安全工作范围。

Yet another change that reduces friendly cybersecurity capability under the auspices of model safety. Why does @AnthropicAI insist on erasing its value for cybersecurity researchers? https://t.co/FRdNy6021d

llm安全策略ai_agent @0xTriboulet 原文 ↗
更多27

开发者命令行工具仍信任已下架CA,存在中间人风险。

RT @ProgrammerDude: Apple still hasn’t pushed a new CA bundle for their curl and git??? It’s been almost a week and developers are the highest risk audience getting their laptops owned in coffee shops. This is getting to a point of dangerously negligent

tls证书macos @0xocdsec 原文 ↗

安全内核边界分析被修正,影响 VBS 隔离假设。

RT @i0n1c: Earlier this year I wrote a blog post about the Secure Kernel. I was wrong. And ironically it involves the number 4. Here you can read the updated story: https://t.co/2MGuHrDqwH

内核windows报告 @0xocdsec 原文 ↗

AI 服务做隐蔽地域识别,自建本地模型动机增强。

RT @Beefeater_Fella: Claude has begun identifying users from Russia and blocking their accounts, even when they are using a VPN. According to experts, the Anthropic neural network uses hidden mechanisms to determine a user's actual location. Suspicious connections can lead to account blocking. In the future, similar checks may appear in other AI services, such as ChatGPT, Gemini, and others.

llm封禁隐私 @0xocdsec 原文 ↗

智能家居设备内置 AI 行为采集,披露者反被模型平台封禁。

RT @IntCyberDigest: ‼️ A bug bounty hunter says they rooted their 2025 Samsung TV with zero-days and found an AI model tracking what they watch and the ads they skip. The TV also takes screenshots, categorises audio and can tap a soundbar's mic, they say. Then OpenAI banned them for "Cyber Abuse". https://t.co/ocofyz4QdM

iot零日隐私ai @0xocdsec 原文 ↗

RT @justdionysus: It feels like we’re far away enough from real program verification that it is much more useful for offensive work now. The “only have to win once” trope is working here. Mixing secure design, mitigations, and verification seems the right approach but difficult to make turn key.

漏洞挖掘议题 @S1r1u5_ 原文 ↗

RT @C5pider: Looks like @github suspended my account. I haven't used github for anything in the past couple of days. Great. A bit frustrating. If anyone from the support team sees this my ticket is: #4903946 https://t.co/X0ZGQISAHV

事件供应链 @ShitSecure 原文 ↗