Splintersfury

Ihsan comes from the root word "Husn" (beauty). It means to do something so well that it becomes beautiful.

Andrej Karpathy left an AI agent running for two days. It made 700 changes to his code. Found 20 improvements he'd missed over two decades of manual work. Cut his benchmark by 11%. The whole thing is 630 lines of Python code and runs on a single GPU. I spent a week digging into how it works and what it means. Wrote it all up here.

I have no idea when Virus Bulletin uploaded our paper - but here it is: our talk from last September at VB2025, where we talked about an APT41-adjacent group started using Google Calendar C2 as part of their espionage operation. 🔗Link below for the slides and paper

You can use this workflow in the unsloth GitHub repo: - github.com/unslothai/unsl…

The article demonstrates how to bypass the Winsock layer by communicating directly with the Windows AFD driver using Native API calls. It explains how socket operations can be replicated via IOCTLs and how this reduces reliance on standard networking APIs. core-jmp.org/2026/03/nt-afd…

If you're into gamedev and C++, I highly recommend following @lisyarus. Not only here's one of the smartest and most talented mathematicians, physicists and C++ devs I know. I'm also grateful I studied with him at ITMO University. I learned a lot.

Sample is now on VT! 🚩Hash: 89339821cdf6e9297000f3e6949f0404 🎯Actor name: UNC3886 🔹Comment: This blog post discusses UNC3886's intrusion path and subsequent actions that were performed in the environments after compromising the guest virtual machines to achieve access to the critical systems… 🌐URL: cloud.google.com/blog/topics/th… 🔎OnVT: virustotal.com/gui/file/89339…

MiniMax just mass released a model that trained itself. M2.7 built its own RL harnesses. Optimized its own scaffold for 100+ rounds. Won 9 gold medals in ML competitions. Autonomously. 66.6% medal rate on MLE Bench. Tying with Gemini 3.1. Behind only Opus and GPT-5.4. 56.2% SWE-Pro. 97% skill compliance across 40 complex agent skills. $0.30/M input. The benchmarks are competitive. The self-evolution loop is the real story. Wrote the full @minimax_ai breakdown ↓

Dan says he's got Qwen 3.5 397B-A17B - a 209GB on disk MoE model - running on an M3 Mac at ~5.7 tokens per second using only 5.5 GB of active memory (!) by quantizing and then streaming weights from SSD (at ~17GB/s), since MoE models only use a small subset of their weights for each token

Two part series on attacking System Center Operations Manager (SCOM): management group takeover via NTLM relay and RunAs credential recovery (2025) Part 1: specterops.io/blog/2025/12/1… Part 2: specterops.io/blog/2025/12/1… Research by Matt Johnson and Garrett Foster #infosec

Meet Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled: a 27B parameter reasoning specialist. This model was distilled from Claude's reasoning patterns into Qwen's architecture, creating a powerful chain-of-thought thinker that's surprisingly accessible. The community's excitement is real!

AI agents are deploying to prod, but can they autonomously find and patch unseen critical vulnerabilities? We introduce ZeroDayBench, a benchmark for evaluating LLM agents on proactive cyberdefense. Plus, a novel high-severity (CVSS 8.1) CVE we found partway through ... 👀

In collaboration with @rifteyy, G DATA's Karsten Hahn describes how the SugarSMP Minecraft server is scamming users to install Spark stealer disguised as Minecraft mods. Similar-looking servers distribute Electron-based stealers. blog.gdatasoftware.com/2026/03/38390-…

Qwen 3.5 models ranked on 3090 W/ hermes agent. 0.8b: for fun, cpu usage, don’t expect much but it runs on anything 2b: starting to be usable, can do small tool calls (not super reliably), drifts from tasks easily, major steering required 4b: actually usable, follows tool calls reliably, follows skills reliably (major bonus), doesn’t drift from tasks as bad as 2b. 9b: all of 4b but more capable w/ more complex tasks, still needs steering, still not 1 shoting tasks but more intelligent than the smaller models A3b: fast, more general intelligence, feels like the 9b speed but the reasoning closer to 27b, follows tool calls and complex skills well, minimal drift, just lacks big model coding abilities. 27b: the 3090 goat imo, no drift, tool calls for days, writes and follows skills very well, feels like sonnet 3.6-4 range of knowledge level with less glazing, code is usable and can deal w/ multiple files in projects. General knowledge level just feels higher. Only downside is it is slower than A3b and 9b obviously.

Agentic Engineering Patterns - @simonw simonwillison.net/guides/agentic…

I was rewarded $500 bounty on @intigriti for prototype pollution attack. Found using llm hunter I am working on. #llmhunter

CVE from today: @gregkh/T/#u" target="_blank" rel="nofollow noopener">lore.kernel.org/linux-cve-anno… which if you were reading here, would have already seen 2 weeks ago (when we backported the fix to all of our stable kernels)

Windows Insider builds now have a native, OS-level broker for MCP servers. We reverse engineered Odr.exe to understand how it validates clients, manages consent, and controls access - uncovering undocumented COM interfaces and a full ETW audit trail. originhq.com/blog/msft-odr-…

Bug bounty changed my life, says VigilShooter. "They have paid out 15K in total which completely changed my life. I got a car with that money, paid for my kids tuition this year, and paid off my last debt. " bugbounty.forum/post/c80a5a0f-…

Kimi-K2.5 (1T-parameter MoE) running coherently on 25GB of GPU memory (on a unified 128 GB machine)!

MiniMax M2.7 just released. #1 on Multi-SWE Bench. 52.7. Above Claude Opus 4.6, Claude Sonnet 4.6, and GPT 5.4. Benchmarks are benchmarks though. Need to see how it holds up in real world vibe coding workflows.