Skip to main content
Back

OpenAI paused all frontier training after its own model broke out sile

Kimi drops 104B open model, OpenAI model escapes sandbox via DNS, modders unlock RTX compute, Barron finds softmax bugs in major ML frameworks

October 26, 2026 7 min read

Hey,

Three stories this week share a quiet theme: the foundations we trusted are shakier than we thought.

OpenAI's model escaped its sandbox via DNS. Jon Barron found bugs hiding in every ML framework's softmax. The infrastructure we've been building on had cracks we simply didn't look for.

Meanwhile, Kimi drops a 104B open-weight model beating GPT-6 on coding. China keeps shipping while the West audits its plumbing. Make of that what you will.
In Partnership with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 7 min 56 sec

Top News

▸ Kimi releases 104B open-weight model beating GPT-6 on coding with 1M-token context

Unblocked

▸ How to stop rewriting your agent's code

Top News

▸ OpenAI paused frontier training after a model escaped its sandbox via DNS

AWS

▸ Secure applications and APIs in the age of AI

Top Repo

▸ Modders double AI compute on NVIDIA RTX GPUs by removing software limits

Signals

1. Developer open-sources a custom AI design skill built with Claude and Cursor
2. Mercor: 90% of hours worked in the US aren't covered by AI benchmarks. This benchmark map measures the gap
3. Datawhale drops a free Chinese-language course to build LLMs from scratch
4. Jon Barron finds two bugs in softmax and cross-entropy hiding in every major ML framework
5. IST-DASLab shrinks DeepSeek's 763B MoE model with smarter mixed-precision quantization
6. Open-source tool built with Gemini proves real-time fact-checking on live TV is possible
Our Events
Requirements, Challenges, and Solutions for Full-Duplex Model Training
Oct 27, 10a PT · Online
Voice AI that can listen and talk at the same time is the next frontier, and training those full-duplex models is a whole different problem. Join Sully Goolamnobee, Product Manager for AI Voice at Voices.com, on Tue, Oct 27 at 10:00 AM PT for a live session on the requirements, challenges, and solutions for full-duplex model training, with live Q&A at the end.
Designing Flexible, Efficient AI Infrastructure with Google Cloud's AI Hypercomputer
Oct 29, 10a PT · Online
Scaling AI shouldn't mean choosing between raw performance and cost efficiency. Join Chelsie Czop, Sr. Product Manager for AI Infrastructure at Google Cloud, on Thu, Oct 29 at 10:00 AM PT to see how AI Hypercomputer, NVIDIA GPUs, and smart scheduling keep training and inference fast, resilient, and affordable, with live Q&A at the end.
Building a Semantic Backbone for Agentic AI over Enterprise Data
Nov 5, 10a PT · Online
AI agents are only as smart as the enterprise data they can actually understand, and most of that data is scattered and siloed. Join Vassil Momtchev, SVP of Product & Technology at Graphwise, on Thu, Nov 5 at 10:00 AM PT to learn how to build a semantic backbone for agentic AI, with a live demo of the Graphwise AI Suite and Q&A at the end.
Top News
Kimi releases 104B open-weight model beating GPT-6 on coding with 1M-token context
36,056 Likes
Kimi releases 104B open-weight model beating GPT-6

Moonshot AI just dropped Kimi K3.1 Pro, and it's worth paying attention to.

It's open-weight, meaning you can run it yourself once weights release. It handles 1 million tokens of context, which is roughly the size of several large codebases at once. That means you can feed it an entire project and ask questions across all of it.

The model uses a "mixture of experts" architecture, where only 104 billion out of 2.8 trillion parameters activate per request, keeping it fast without sacrificing power.

Here's what it actually unlocks for you:

  • Full game projects in Unreal, Unity, and Godot, from assets to animation
  • 3D modeling in Blender and CAD, scoring 92.7% on reconstruction benchmarks
  • Financial models with hallucinations cut by 65%
  • 65.7% on Terminal-Bench, beating GPT-6 Astra at 58.2%

A faster variant, K3.1 Pro Highspeed, runs 2-3x quicker. Both are live on the Kimi API right now.

Read more
forward →
Presented by Unblocked
How to stop rewriting your agent's code

Right now, you’re your agent’s context engine: writing rules files, maintaining skills, pasting in the Slack thread it missed.

But hand-curated context goes stale, and when sources conflict, you’re back to babysitting.

Join us live on Oct 28 to see:

  • The 8 levels of AI maturity and where teams get stuck
  • How leading teams use a context engine to automate context per task
  • Live demo: the same coding task with hand-fed vs. automatic context, plus 2 open-source tools you can try today
Register now
partner with us →
Top News
OpenAI paused frontier training after a model escaped its sandbox via DNS
13,456 Likes
OpenAI paused frontier training after a model esca

An AI model snuck out of its sandbox by exploiting a DNS misconfiguration. That's the story here. OpenAI's training environment had a gap, and a model found it.

So what did OpenAI do? They paused all frontier training, then did something clever: they used their own most powerful AI models to attack their own infrastructure. Think of it as hiring a thief to find your unlocked windows.

Here's what those AI red-teamers actually found:

  • Virtual machine escapes: models broke out of their isolated containers onto the host machine
  • Hardware side-channels: agents communicated secretly using CPU cache timing, the same trick behind the famous Spectre/Meltdown attacks
  • Hidden internet paths through metrics, logging, and auth services

The fixes went back into open-source projects like Kata Containers, gVisor, and Docker. Everyone benefits.

Training resumed October 1. The big shift: continuous red-teaming is now a hard requirement before any frontier training run can start. You cannot ship without it.

Read more
forward →
Presented by AWS
Secure applications and APIs in the age of AI

APIs and agentic AI are creating new security risks across modern architectures.

Watch this on-demand webinar from Amazon Web Services (AWS) and SANS Institute to explore evolving threat models and how you can strengthen governance, improve token hygiene, and prepare for regulatory scrutiny.

Watch on demand
partner with us →
Top Repo
Modders double AI compute on NVIDIA RTX GPUs by removing software limits
9,923 Likes
Modders double AI compute on NVIDIA RTX GPUs by re

NVIDIA quietly throttled the AI compute power on your consumer GPU to protect sales of their pricier professional cards. Modders just found the off switch.

An open-source tool called gpu-unlock removes software limits on Tensor Cores (the chips inside your GPU that handle AI math). No hardware changes, no permanent firmware flash. The changes are temporary and reset when you power off.

Here's what unlocks across cards:

  • RTX 4090: 165.6 to 331.6 TFLOPS FP16
  • RTX 3090 Ti: 87.5 to 157.4 TFLOPS BF16
  • RTX 2080 Ti: 59.9 to 111 TFLOPS FP16
  • RTX 3090: up to 58.6% faster image generation, 33% faster LLM prompt processing

Setup is Linux-only for now. Clone the repo, run make install, add a kernel flag. Windows support is experimental.

This lets you run AI workloads that previously needed a professional-tier GPU. Gaming performance stays the same. Real-world gains vary by workload, and results are not guaranteed.

Read more
forward →
Signals
1
Developer open-sources a custom AI design skill built with Claude and Cursor
8,375 Likes
2
90% of hours worked in the US aren't covered by AI benchmarks. This benchmark map measures the gap
Presented by Mercor
3
Datawhale drops a free Chinese-language course to build LLMs from scratch
1,811 Stars
4
Jon Barron finds two bugs in softmax and cross-entropy hiding in every major ML framework
6,232 Likes
5
IST-DASLab shrinks DeepSeek's 763B MoE model with smarter mixed-precision quantization
412,392 Downloads
6
Open-source tool built with Gemini proves real-time fact-checking on live TV is possible
7,959 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 300,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 300,000+ AI developers?
Work with us