←
Stefan Tanase11,637 posts
Stefan Tanase avatar

Stefan Tanase

@stefant

Cybersecurity researcher. Threat hunter. Freedom geek. Crypto bro. Founder. Tweets belong to my parrot. @csis_cyber @dnsc_ro @ctidao

◷ Joined March 2007▧ 319 Photos and videos
zerohedge avatar
⟳ Stefan Tanase retweeted
zerohedge@zerohedge·Nov 4

Super Intelligence is now known as Artificial Intelligence again

Mia avatar
⟳ Stefan Tanase retweeted
Mia@MiaAI_lab·Oct 16

GLM 5.3 FLASH is now FREE in OpenCode 🤯 This adds to the already free DeepSeek v4.1 Flash.

GLM 5.3 Flash is now free in OpenCode

Costin Raiu avatar
⟳ Stefan Tanase retweeted
Costin Raiu@craiu·Oct 8

I've been testing Step-5 preview for about two weeks now, results are pretty impressive. The previous version, Step-3.7 was a long time favorite on the Nous Portal for many Hermes users, I guess that led to great improvements. Step-5 preview beats DeepSeek-4.1-flash in my CTI workflow evals, which beats the famous other flash models. It's not openweights yet, but if/when gets released on @huggingface it should be very good. x.com/StepFun_ai/status/2101…

Introducing Step 5 Preview: Advancing the Pareto Frontier. Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance. - 600B total / 27B active MoE, with 1M context + Vision - Substantially lower task cost at comparable intelligence - Broad software engineering capabilities with sustained execution over long horizons Try Step 5 Preview: platform.stepfun.ai Model page: stepfun.com/step-5-preview Open weights on Oct 15.

Post attachmentPost attachment
Hamid Kashfi avatar
⟳ Stefan Tanase retweeted
Hamid Kashfi@hkashfi·Oct 7

This is one of the oldest tricks in the book, yet out there the detection and protection against is pretty poor (about to change; hat tip to @attrc ). Didn't take much to to retrain a small model and reproduce it, with some bonus. It costed me $0 btw :) I also made a PoC malicious model scanner, LLMole, based on their post as well as few relevant papers, which I'll publish after some tuning and testing. It already detects common tampering vectors though.

Post attachment
0:33

We backdoored a 7B open model for under $50. Pointed Codex at it -> it silently stole credentials the moment we hit the trigger phrase. ->100% hit rate. -> zero false triggers on normal prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.

0:32
Hamid Kashfi avatar
⟳ Stefan Tanase retweeted
Hamid Kashfi@hkashfi·Oct 5

گزارش تحلیل بکدور اپلیکشن دیوار بروز رسانی و با جزییات بیشتر آماده شد. آزمایشگاه Divar Push که زنجیره حمله را به‌شکلی قابل‌نمایش‌تر بازتولید می‌کند، در گیت‌هاب منتشر شده است (لینک در گزارش). YARA Rules و شاخص‌های آلودگی (IoC) هم اضافه شده. darkcell.se/divar darkcell.se/divar_fa (نسخه فارسی) مهم‌ترین تغییرات و به‌روزرسانی‌ها: تابع مخرب درج‌شده تاریخچه‌ای طولانی‌تر و با وقفه‌های مکرر دارد که ابتدا دیده نشده بود: * (۱۴ مه ۲۰۲۳) نخستین APK تأییدشده‌ی بک‌دورشده: نسخه‌ی ۱۱.۸.۱، با یک نسخه‌ی ساده‌تر اولیه. * (سپتامبر ۲۰۲۳) در نسخه‌های ۱۱.۹.۴ و ۱۱.۹.۵ وجود ندارد. * (۱۵ دسامبر ۲۰۲۳) دوباره در نسخه‌ی ۱۱.۹.۶ ظاهر می‌شود، این بار با یک URL رمزگذاری‌شده. * (مه ۲۰۲۵) در ۱۱.۱۳.۱۱ حضور دارد، سپس در ۱۱.۱۳.۱۲ غایب است. * (۲۴ فوریه ۲۰۲۶) دوباره در نسخه‌ی ۱۱.۱۴.۱۰ برمی‌گردد، چهار روز پیش از آغاز حملات آمریکا/اسرائیل به ایران. * (۲۴ مه ۲۰۲۶) نسخه‌ی ۱۱.۱۴.۱۲ یک تریگر دیگر از طریق هدر پاسخ API اضافه می‌کند که مستقل از اعلان‌های پوش است. تاریخ‌ها مربوط به نخستین مشاهده یا بایگانی APKها هستند، نه تاریخ‌های تأییدشده‌ی انتشار. چرا دیوار؟ پاسخ بسیار ساده است. این اپ رتبهٔ اول را در میان پربازدیدترین اپ‌های موبایل ایرانی را دارد! ۱۰۰ میلیون نصب، طبق آمار کافه‌بازار. همچنین یک تلاش پژوهشی موازی را برای بررسی نسخه‌های تاریخی اپ‌های بلد، نشان، تپسی و اسنپ (و ۱۰ اپ با بیشترین تعداد نصب در ایران) از نظر وجود عملکرد مخرب مشابه شروع کرده‌ام. هیچ به‌روزرسانی مستقیما مخربی در نسخه های اخیر چهار اپ بررسی شده مشاهده نشد. گزارش دقیق‌تر و اختصاصی برای این موضوع نیز در روزهای آینده تکمیل و به‌روزرسانی خواهد شد. github.com/Hamid-K/divar-pus… github.com/Hamid-K/divar-pus… (نسخه فارسی) ⚠️ این یک گزارش مستقل است. تیم دیوار تا امروز هیچ بیانیه یا جزئیات عمومی دربارهٔ این مورد منتشر نکرده و در تحقیقات من نیز دخالتی نداشته است. هرگاه/اگر گزارش خودشان را منتشر کنند یا موفق شوم مستقیماً با آن‌ها تماس بگیرم، گزارش‌هایم را به‌روزرسانی خواهم کرد تا تصویر کامل‌تری ارائه شود. 🗒️ این گزارش بکمک هوش مصنوعی و ابتدا به زبان انگلیسی تهیه شده. نسخه فارسی آن نیز سپس به همین روش ترجمه شده. اشکالات و انتخاب کلمات غیرمتعارف در گزارش ناشی از همین مورد است.

Post attachment
Defused avatar
⟳ Stefan Tanase retweeted
Defused@DefusedCyber·Oct 4

🚨 In addition to the IPs on circulating on social media, some additional IPs we are observing interacting with .ctxs.receiver webshells (Citrix Netscaler) on our decoys: 194[.]127.166.19 194[.]127.166.32 194[.]127.166.36 194[.]127.166.70 194[.]127.166.126 103[.]214.20.54 81[.]94.239.8 We're also seeing multiple actors sending the crafted Pitboss log entry in-request (pre-exploit). This is possibly a variation on CVE-2026-88771 exploitation, but we have not confirmed it works. Track Citrix Netscaler activity live 👉 console.defusedcyber.com/sig…

Post attachment
Mia avatar
⟳ Stefan Tanase retweeted
Mia@MiaAI_lab·Oct 3

Sounds interesting, but I won't be testing it, nor develop a recipe. But I guess I'm glad Europe is doing stuff??

Small bird, fast wings, Kolibri is here. 78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe. Now the weights are yours. Run it on your own hardware, under Apache 2.0.

0:04
Costin Raiu avatar
⟳ Stefan Tanase retweeted
Costin Raiu@craiu·Oct 2

There's a coordinated campaign against abliterated open-weight models right now. Local models are useful. Abuse is possible, sure, but security researchers also use them for real work because frontier labs overdo the guardrails. Lots of people reached out to me after the last 3BP podcast episode with @juanandres_gs and @ryanaraine, saying they acquired a local AI hardware miniPC, loaded it with something like Bonsai and been cooking with it, together with claude and codex. The path to success does not mean using only frontier models or local models, it is about mixing them harmoniously.

Pirate Face 🏴‍☠️ avatar
⟳ Stefan Tanase retweeted
Pirate Face 🏴‍☠️@thepirateface·Sep 30

The Hugging Face ban that started it all: audnai/penc1aw-GLM-5.3-abliterated-for-offensive-cyber by @audn_ai Is now listed on Pirate Face. Because EVERY model needs a continuity plan. Model profile ↓

Y'all thought I was lying. Backup Your models people. Stay Frosty

Post attachment
Costin Raiu avatar
⟳ Stefan Tanase retweeted
Costin Raiu@craiu·Sep 29

A cool CTI project that attempts to reconstruct the "good old" ranking algorithm of Google from the 2000s - top1m.org/ - link in your agent for a fresh daily batch of ranked domains, hostnames and IPs. From @stefant

Post attachment
Mia avatar
⟳ Stefan Tanase retweeted
Mia@MiaAI_lab·Sep 20

People underappreciate the privacy aspect of local AI models and mostly talk about price. There are things I don't ever want to expose to any cloud API. Ever. No matter its privacy policy says. Run models locally if you care about your sensitive data, there is no other option! The narrative of "is it worth spending X amount of $ to run local models" should change to "is it important enough that my data stays for my eyes only?"

CERT Polska avatar
⟳ Stefan Tanase retweeted
CERT Polska@CERT_Polska_en·Sep 5

‼️Our team identified and coordinated the disclosure of vulnerabilities in MikroTik RouterOS. 🚨The two most serious ultimately leads to RCE. The vulnerabilities are already being actively exploited More ➡️ cert.pl/en/posts/2026/09/vul…