
We're excited to be at #POC2026 for the first time! Come up to our booth and say hi, we are hiring!


We're excited to be at #POC2026 for the first time! Come up to our booth and say hi, we are hiring!

Will be at #POC2026 next week! Hit me up at the @catalystsec booth and come chat about AI, bug hunting, and anything in between. Also we are hiring :)
Thank you for sponsoring #POC2026 as a GOLD! Catalyst Security (@catalystsec) is a growing team of highly experienced vulnerability researchers, focused on solving the most challenging technical problems. Come and meet us!

Thank you for sponsoring #POC2026 as a GOLD! Catalyst Security (@catalystsec) is a growing team of highly experienced vulnerability researchers, focused on solving the most challenging technical problems. Come and meet us!

Two thoughts on this: 1. Very cool. As someone who's primary interest is the offensive side of cyber, I love it. *chefs kiss* and congrats to the team. 2. You cannot bug hunt your way to 'secure'. AI doesn't change that. If your strategy for the next few years is to boil the oceans continuously scanning your same codebase, with the same broken architecture, with the same broken threat models, you are going to have a bad time. a) you're going to spend a fortune and b) you're going to keep getting hacked.
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment. The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work. GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time. openai.com/index/introducing…

Some nice speed improvements for DeepSeek V4.1 Flash got merged to DS4! No better time to run this model at native precision on your Mac 😃 github.com/antirez/ds4/pull/…

Is this the first time Meta and Apple work together on CEO level? MLX is the preferred local inference platform for Meta, and we get 0-day support for Muse Spark Open in mlx-lm with vision support. A lot of people use third party inference engines other than mlx-lm, and they will all benefit from the optimization works for Muse Spark Open in MLX. Which is an important meaning of the mlx-lm project: to provide a good foundation for supporting new models, not to compete with open source community. For the technical details, our new blog is a great read: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio machinelearning.apple.com/re…
To help make advanced models more accessible, Muse Spark Open is optimized to run on two Apple M5 Ultra Mac Studios with the latest MLX framework. It also runs on leading inference engines like SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Developers can further customize the model with PyTorch, including TorchTitan for fine-tuning.

Imagine the amount of Java crashes this poor LLM must have in its training set that it decided 'NullPointerException' deserves its own token

When you can decompile all the things
MW2 decompilation is nearing its end. With currently 17 agents working in parallel, we're making progress at an incredible pace. The game is fully playable now and almost all visible bugs are gone. Only a few are left. I even managed to get my first nuke ☢️😂

FYI, this turned out to be a massive performance regression for MiMo models and I imagine GLM/Qwen may exhibit similar characteristics.

If you're using DS4 for serving GLM or Qwen models, you'd want this one: github.com/antirez/ds4/pull/…

Xiaomi published their RL training stack for MiMo-V2.6. Reveals its cyber RL primarily targets vulnerability reproduction (using ARVO/OSS-Fuzz), not specifically exploitation. huggingface.co/datasets/Xiao…

Try comparing the Windows kernel to XNU 😆
Say what you will about Windows (OS), but the NT kernel really is an engineering marvel that still puts Linux to shame in many ways. The quickest way to describe it for a programmer, is NT was more like an object-oriented language, with a strong security model from day one, whereas Linux is very…not. A lot of the “good” features in Linux feel bolted-on (SELinux, Capabilities, Namespaces) because…well they were. I love to imagine an alternate history where NT won. IMO, Microsoft *should* have made an “Open NT” in the early 2000s; not fully GPL-style open, but one where a large org could say…swap out a memory allocator for their own. (they sorta did this with limited source access, but it was too restrictive) You could imagine say…an early Amazon forking OpenNT to create an “AmazonNT” for EC2, where they have a modified scheduler, network stack, whatever. But, the security+compatibility contract keeps a stable baseline on the Microsoft side. Controversial take, but if we enter this era where users are giving AI agents increasingly higher levels of access control; the Linux kernel is legitimately a poor fit. Think about it; answer the question “What exactly is this AI agent allowed to do?” On Standard Linux, it’s disgustingly messy with lots of overlap. Do you use UIDs? GIDs? ACLs? CGROUPs? Policies? SELinux? Filesystem modes? There’s not a singular coherent graph of capabilities you can point to. Too many ways you can escape an initially narrow scope. NT, by comparison, can go the route of explicitly typed resources, and then you could have these really strong centralized audit trails when an agent goes haywire…etc. I know I’m rambling, but the point is…if you were greenfielding an OS kernel from scratch, in 2026, with the intent of being forward-looking, it would *not* look like Linux. Frankly, it’d probably look a lot closer to NT, or even a BSD fork…

Comparing MiMo-V2.6-Flash in DS4 to the Xiaomi API. More work to be done on long context :) Still cool that a single M3 Ultra can compete on decode performance. Also, prefill on M5 Ultra should be much better.

Btw, there was heavy focus on accuracy for this work, so no shortcuts with kernels that have rounding errors etc. Hence, long context currently pays a penalty, but should be salvageable with some additional effort.

I put my @UnpromptedAU slides up at justdionysus.github.io/slide… — a bit of reflection on exploit development in the age of AI. My TL;DR is keep pushing to understand complex things, be honest with your own understanding, and use AI as a power tool to increase pace and depth.
