# Matt Suiche

> Hacker · Founder of OnDB. Personal site of Matt Suiche — security research, AI/ML systems work, and notes from building things.

This file follows the llms.txt convention (https://llmstxt.org). Every post is also available as raw Markdown by appending `index.md` to its URL. The full corpus is at /llms-full.txt.

## About

- [About Matt Suiche](https://www.msuiche.com/about/): background, current focus, contact

## Posts

- [The 1.2 ms Eigensolver That Never Ran](https://www.msuiche.com/posts/1-2ms-eigensolver-that-never-ran/): A real batched symmetric eigensolver reached 29.9 ms on an NVIDIA B200. A cache probe scored 1.203 ms by moving all numerical work outside the timer and still passed every check. Here is the solver, the harness flaw, and the corrected measurement protocol.
- [A Windows Kernel in a Browser Tab, Part III: Debugging It, and the Crash Dumps It Writes Itself](https://www.msuiche.com/posts/nanokrnl-debugging-crash-dump/): You can attach lldb to nanokrnl while it runs in a browser tab, break in kernel code, and hit it. When it bugchecks it writes two crash dumps of itself: a Linux ELF core symbolized by its own DWARF, and a native Windows MEMORY.DMP that WinDbg opens as a full kernel target, with lm, dt, r, kv, !analyze -v, !object and !process all working. From a kernel that has never run on real hardware.
- [A Windows Kernel in a Browser Tab: Running Unmodified Microsoft Console Tools](https://www.msuiche.com/posts/nanokrnl-windows-console-tools/): The prompt you type into at nanokrnl.ai runs the real Microsoft cmd.exe, sort.exe, where.exe, whoami.exe and more.com, unmodified, on a from-scratch Rust NT kernel's own syscalls. Here is what it takes to load a real .exe, give it handles, and let it spawn children, plus the honest frontier: pipes.
- [A Windows Kernel in a Browser Tab, Part II: A Filesystem Over 9P, From a JavaScript Object](https://www.msuiche.com/posts/nanokrnl-9p-host-filesystem/): nanokrnl has no disk. It runs in a browser tab. So how does 'more H:\readme.txt' read a real file? The answer is 9P, the Plan 9 protocol that Linux and WSL2 use to share a host filesystem into a guest. This post walks the transport, the kernel client, the in-page JavaScript server, and one very confusing bug on the way to a working H: drive.
- [A Windows Kernel in a Browser Tab, Part I: Cold Boot, Fast Boot, and Four Megabytes](https://www.msuiche.com/posts/nanokrnl-cold-boot-fast-boot/): nanokrnl is an NT-shaped kernel in Rust that boots in a browser tab through nanox, a 65 KB x86-64 emulator we wrote from scratch. This first post in a nanokrnl and nanox series walks the two ways it boots, how the whole machine is snapshotted and restored, and why the running system fits in about four megabytes when a modern OS asks for gigabytes.
- [Fable 5's 38-Minute Kernel, Part II: The Token Math and the Boot Count](https://www.msuiche.com/posts/fable-5-windows-kernel-part-2/): Part I said Fable 5 wrote a booting NT-shaped kernel in 38 minutes. The viral reply asked the right question: can a model output that many tokens, or compile and boot often enough, in 38 minutes? Part II answers it from the transcript, and the answer inverts the skepticism.
- [CVE-2010-2568: Stuxnet's .LNK Zero-Day, Line by Line in the Windows 2000 Source (GLM-5.2 Analysis)](https://www.msuiche.com/posts/cve-2010-2568-stuxnet-lnk/): CVE-2010-2568 let Stuxnet execute code the instant a user opened a folder. Read from the leaked Windows 2000 source, it is not a memory bug, and that is why it would have compiled and shipped under Rust too.
- [Fable 5 wrote a Windows kernel in 38 minutes](https://www.msuiche.com/posts/fable-5-windows-kernel/): Fable 5 wrote a booting, NT-shaped kernel in Rust in 38 minutes. The code is impressive; the signal is what it tells us about AI-authored trust, and the verification frontier that decides whether the internet's critical infrastructure gets rewritten by models.
- [Windows 11 Hibernation on ARM64: the Boot Manager, winresume, and the hiberfil.sys Format](https://www.msuiche.com/posts/windows-11-arm64-hibernation/): A full reverse-engineering of the Windows 11 25H2 ARM64 hibernation and resume path: how the boot manager decides to resume, how winresume.efi parses and restores hiberfil.sys, the on-disk PO_MEMORY_IMAGE header and compressed-block format, the encryption model (BitLocker for the bulk, AES-GCM for the secure-kernel section), the TRIM-on-resume behavior, and the ARM64 processor-state restore including the PAC keys and SVE. With what is new in Windows 11 and what is specific to ARM64.
- [SMBaloo, Part II: An AI Agent, the ARM64 Genericity Gap, and Windows 11 Kernel Internals](https://www.msuiche.com/posts/smbaloo-part-ii-an-ai-agent-the-arm64-genericity-gap-and-windows-11-kernel-internals/): People keep claiming AI agents can discover new exploitation techniques. I gave Twinkle a falsifiable version: take SMBaloo, the 2020 ARM64 SMBGhost exploit Matt admitted he never made fully generic, find the one step that pins it to specific hardware, and close it. Then reverse the boot manager, OS loader, and kernel from a current Windows 11 25H2 ARM64 ISO to verify the memory management: the deterministic page-table allocator, the fixed 0xFFFFF68000000000 self-map still live in the shipping kernel, and the GIC locate target surviving six years into a PAC-hardened build.
- [autoextdetector: A Self-Improving Detection Agent for Supply-Chain Attacks](https://www.msuiche.com/posts/autoextdetector-a-self-improving-detection-agent-for-supply-chain-attacks/): The 14 detectors and ~62 rules aren't the interesting part. The interesting part is that the same loop that built them is the loop that repairs them when production surfaces a false positive — usually within an hour, costing under a dollar. Bounded recursive self-improvement in the auto-research lineage Karpathy has been describing: agent proposes, sandboxed validator verifies, deterministic Pareto gate decides keep-or-discard, journal records every attempt. Two days ago I published the cluster analysis of 230,000 OSV advisories; today I'm open-sourcing the detection agent. Zero ML at runtime; bounded RSI offline. Here's how the loop closes.
- [Supply-Chain Attacks Cluster: 230,000 Advisories, Five Patterns](https://www.msuiche.com/posts/supply-chain-attacks-cluster-230000-advisories-five-patterns/): Pulled the full OSV mirror for npm and PyPI — 230,000+ advisories. The malicious-tagged subset clusters into five recurring patterns. None of them are clever. All of them keep working. A note on why two decades of EDR/XDR investment is structurally unable to stop the next event-stream.
- [From Y2K to Patch Tuesday 2025: 25 Years of Bugs in the Windows 2000 Source Tree](https://www.msuiche.com/posts/from-y2k-to-patch-tuesday-2025-25-years-of-bugs-in-the-windows-2000-source-tree/): A scoped audit of the leaked Windows 2000 source tree turned up 45 findings mapped against 24 public CVEs spanning 1999 to 2025 — including four bugs that took Microsoft between 6 and 25 years to patch, plus the WMI EoPs R00tkitSMM found by auditing the leaked Windows Research Kernel source.
- [Bleeding Llama: When AI Model Files Become Memory Leaks](https://www.msuiche.com/posts/bleeding-llama-when-ai-model-files-become-memory-leaks/): Technical analysis of CVE-2026-7482, a critical unauthenticated heap OOB read vulnerability in Ollama's GGUF processing that leaks process memory through quantization
- [Legacy Security Is the Real Enterprise AI Bottleneck](https://www.msuiche.com/posts/legacy-security-is-the-real-enterprise-ai-bottleneck/): Security, not model capability, is what's blocking Agentic AI in the enterprise, where the real market is. MCP misuse, the Lovable and Mercor breaches, the Vercel/Context incident, Delve's SOC 2 problems: the AI ecosystem is failing at fifteen-year-old bug classes while talking about AGI, and compliance theater is not going to fix it.
- [Seeing Sound: Generative Techno and DSP in Pure NumPy](https://www.msuiche.com/posts/seeing-sound-generative-techno-and-dsp-in-pure-numpy/): Building a complete generative techno engine and a reference-track analyzer in pure NumPy — from raw oscillators to 8 arranged-track genre presets. Covers DSP fundamentals, envelopes, LFOs, filters, sidechain, generative patterns, arrangement, reverse-engineering real tracks, and seeing sound as frequency bands.
- [From La Fontaine to Lego: Characters as Ideological Delivery Systems](https://www.msuiche.com/posts/from-la-fontaine-to-lego-characters-as-ideological-delivery-systems/): From La Fontaine's animal fables to USAID's Muppets to Iran's GenAI Lego rap videos — the use of characters as ideological delivery systems is centuries old. AI didn't invent the playbook. It made it cheaper, faster, and deeper.
- [Local Models Within Reach: Everything That Changed in Eight Months](https://www.msuiche.com/posts/local-models-within-reach-everything-that-changed-in-eight-months/): A follow-up to my August 2025 notes on building agents for small language models. Gemma 4, Qwen3.5 MoE, TurboQuant, MLX, and a memory market correction have quietly made local AI the default option.
- [Odd Lots: Cyberwar in the Age of AI](https://www.msuiche.com/posts/odd-lots-cyberwar-in-the-age-of-ai/): Notes from my second appearance on Bloomberg's Odd Lots podcast -- cyberwar, Iran, the AWS datacenter strikes, Anthropic vs. the Pentagon, and why software is going to zero.
- [When Machines Pay Machines: The Economics of Agentic AI](https://www.msuiche.com/posts/when-machines-pay-machines-the-economics-of-agentic-ai/)
- [Porting CUDA FFT to Mojo: Achieving Bit-Exact Precision](https://www.msuiche.com/posts/porting-cuda-fft-to-mojo-achieving-bit-exact-precision/): Porting a CUDA Fast Fourier Transform implementation to Mojo required deep PTX assembly analysis and understanding floating-point nondeterminism to achieve bit-exact precision matching
- [AMD GPU Support in Triton Gluon Framework](https://www.msuiche.com/posts/amd-gpu-support-in-triton-gluon-framework/): Technical analysis of AMD GPU support implementation in Triton's Gluon framework, including architecture-specific optimizations and performance characteristics.
- [RustBPE: High-Performance BPE Tokenizer Training in Rust](https://www.msuiche.com/posts/rustbpe-high-performance-bpe-tokenizer-training-in-rust/): Analysis of RustBPE - a Rust implementation of BPE tokenizer training with parallel processing and performance optimizations over Python implementations.
- [Optimizing AlphaFold's Triangle Multiplicative Update: A First Look at GPU Performance Engineering](https://www.msuiche.com/posts/optimizing-alphafolds-triangle-multiplicative-update-a-first-look-at-gpu-performance-engineering/): Learning GPU performance engineering through the GPU MODE TriMul challenge - achieving 2.42× speedup on H100 through FP16 optimization, weight fusion, and systematic experimentation.
- [Multi-GPU Programming with AMD's Iris Framework for Triton](https://www.msuiche.com/posts/multi-gpu-programming-with-amds-iris-framework-for-triton/)
- [Gluon: When Triton Isn't Low-Level Enough](https://www.msuiche.com/posts/gluon-when-triton-isnt-low-level-enough/): After diving deep into PyTorch, Triton, CUDA, and PTX, I discovered Gluon - Triton's answer to the performance gap that even Triton can't always bridge. Here's what I learned.
- [CVE-2025-21043: When DNG Opcodes Become Attack Vectors](https://www.msuiche.com/posts/cve-2025-21043-when-dng-opcodes-become-attack-vectors/)
- [The Hidden Math Bug That Makes AI Unpredictable](https://www.msuiche.com/posts/the-hidden-math-bug-that-makes-ai-unpredictable/)
- [Building Agents for Small Language Models: A Deep Dive into Lightweight AI](https://www.msuiche.com/posts/building-agents-for-small-language-models-a-deep-dive-into-lightweight-ai/): Exploring the architecture, challenges, and implementation patterns for building AI agents with small language models (270M-32B parameters) that can run on consumer hardware
- [ELEGANTBOUNCER: When You Can't Get the Samples but Still Need to Catch the Threat](https://www.msuiche.com/posts/elegantbouncer-when-you-cant-get-the-samples-but-still-need-to-catch-the-threat/): The story of how ELEGANTBOUNCER was born from the frustration of not having access to in-the-wild exploit samples, and why structural analysis beats signatures for advanced mobile threats
- [Detecting CVE-2025-43300: A Deep Dive into Apple's DNG Processing Vulnerability](https://www.msuiche.com/posts/detecting-cve-2025-43300-a-deep-dive-into-apples-dng-processing-vulnerability/): Technical analysis and detection methodology for CVE-2025-43300, a critical 0-click RCE vulnerability in Apple's DNG image processing
- [Bob and Alice in Kernel-land - Part 3](https://www.msuiche.com/posts/bob-and-alice-in-kernel-land-part-3/)
- [Bob and Alice in Kernel-land - Part 2](https://www.msuiche.com/posts/bob-and-alice-in-kernel-land-part-2/)
- [Financial Forensics in a fragmented ecosystem](https://www.msuiche.com/posts/financial-forensics-in-a-fragmented-ecosystem/)
- [Election Security - Friday Review](https://www.msuiche.com/posts/election-security-friday-review/)
- [Bob and Alice in Kernel-land](https://www.msuiche.com/posts/bob-and-alice-in-kernel-land/)
- [Researching Triangulation: Detecting CVE-2023-41990 with single byte signatures.](https://www.msuiche.com/posts/researching-triangulation-detecting-cve-2023-41990-with-single-byte-signatures./)
- [Researching BLASTPASS: Analysing the Apple & Google WebP POC file - Part 2](https://www.msuiche.com/posts/researching-blastpass-analysing-the-apple-google-webp-poc-file-part-2/)
- [Researching BLASTPASS: Detecting the exploit inside a WebP file - Part 1](https://www.msuiche.com/posts/researching-blastpass-detecting-the-exploit-inside-a-webp-file-part-1/)
- [Researching FORCEDENTRY: Detecting the Exploit With No Samples](https://www.msuiche.com/posts/researching-forcedentry-detecting-the-exploit-with-no-samples/)
- [POC 2022 - Korea - Keynote 🦀](https://www.msuiche.com/posts/poc-2022-korea-keynote/)
- [Vegas 2022 - A web3 security review](https://www.msuiche.com/posts/vegas-2022-a-web3-security-review/)
- [Magnet Forensics Acquires Cybersecurity Software Firm Comae Technologies ](https://www.msuiche.com/posts/magnet-forensics-acquires-cybersecurity-software-firm-comae-technologies/)
- [SUNBURST & Memory Analysis](https://www.msuiche.com/posts/sunburst-memory-analysis/)
- [Azure Sphere Internals - Overview](https://www.msuiche.com/posts/azure-sphere-internals-overview/)
- [SMBaloo - Building a RCE exploit for Windows ARM64 (SMBGhost Edition)](https://www.msuiche.com/posts/smbaloo-building-a-rce-exploit-for-windows-arm64-smbghost-edition/)
- [Twitter's Information Operations - An OSINT Analysis](https://www.msuiche.com/posts/twitters-information-operations-an-osint-analysis/)
- [Facebook's Coordinated Inauthentic Behavior - An OSINT Analysis](https://www.msuiche.com/posts/facebooks-coordinated-inauthentic-behavior-an-osint-analysis/)
- [How to Solve the Blindspots of Event-Driven Detection](https://www.msuiche.com/posts/how-to-solve-the-blindspots-of-event-driven-detection/)
- [Rethinking Logging for Critical Assets](https://www.msuiche.com/posts/rethinking-logging-for-critical-assets/)
- [Smart Contract Languages to Follow](https://www.msuiche.com/posts/smart-contract-languages-to-follow/)
- [Porosity: A Decompiler For Blockchain-Based Smart Contracts Bytecode](https://www.msuiche.com/posts/porosity-a-decompiler-for-blockchain-based-smart-contracts-bytecode/)
- [Petya.2017 is a wiper not a ransomware](https://www.msuiche.com/posts/petya.2017-is-a-wiper-not-a-ransomware/)
- [Petya— Enhanced WannaCry ?](https://www.msuiche.com/posts/petya-enhanced-wannacry/)
- [Lessons from TV5Monde 2015 Hack](https://www.msuiche.com/posts/lessons-from-tv5monde-2015-hack/)
- [WannaCry — Decrypting files with WanaKiwi + Demos](https://www.msuiche.com/posts/wannacry-decrypting-files-with-wanakiwi--demos/)
- [WannaCry — Links to Lazarus Group](https://www.msuiche.com/posts/wannacry-links-to-lazarus-group/)
- [WannaCry — New Variants Detected!](https://www.msuiche.com/posts/wannacry-new-variants-detected/)
- [WannaCry — The largest ransom-ware infection in History](https://www.msuiche.com/posts/wannacry-the-largest-ransom-ware-infection-in-history/)
- [PASSFREELY: Oracle & SWIFT at risk](https://www.msuiche.com/posts/passfreely-oracle-swift-at-risk/)
- [ShadowBrokers: The NSA compromised the SWIFT Network](https://www.msuiche.com/posts/shadowbrokers-the-nsa-compromised-the-swift-network/)
- [Windows 7 and Windows Server 2008 R2 djoin (Offline Domain Join) utility.](https://www.msuiche.com/posts/windows-7-and-windows-server-2008-r2-djoin-offline-domain-join-utility./)
- [Retrieving MmPhysicalMemoryBlock regardless of the NT version](https://www.msuiche.com/posts/retrieving-mmphysicalmemoryblock-regardless-of-the-nt-version/)
- [Check your system virginity in less than 60 seconds.](https://www.msuiche.com/posts/check-your-system-virginity-in-less-than-60-seconds./)
- [X-Ways Forensics Beta 2 and hibernation file. (coincidence?)](https://www.msuiche.com/posts/x-ways-forensics-beta-2-and-hibernation-file.-coincidence/)

## Other

- [Press](https://www.msuiche.com/press/): press coverage and mentions
- [Media](https://www.msuiche.com/media/): media appearances and interviews
- [RSS feed](https://www.msuiche.com/index.xml)
- [Full corpus (Markdown)](https://www.msuiche.com/llms-full.txt)