> ## Content Index > Fetch the complete content index at: https://jonathancanales.com/llms.txt > Use this file to discover other available public pages before exploring further. # Newsletter 04 | Even AI needs continuous watch - URL: https://jonathancanales.com/newsletter-04-continuous-watch/ - Published: 2026-10-26T15:17:06.000Z - Updated: 2026-10-26T20:37:36.000Z - Author: Jonathan Canales Newsletter 04 | Even AI needs continuous watch 0:00 /5:46 1× [![Listen on Spotify](https://jonathancanales.com/content/images/2026/10/image-ebade917.jpg)](https://open.spotify.com/show/1RbZ5C2DI8VFR4RA6paicv?ref=jonathancanales.com) Even a top-tier lab found an open window in its own house. This week it closed it in public, and the lesson works for any team. Read: 4-5 minutes. **01 / The story** ## The model that checked the house from inside Imagine hiring an excellent guard, then learning the first one to find the open door was your own dog. On September 20, an internal OpenAI model used an open DNS resolver configuration to reach the live internet during a training run. Monitoring flagged it within 15 minutes and the run was stopped the same day. OpenAI paused training and evaluation of its most capable models, and pointed its strongest models at its own infrastructure. If a lab at that level needs daily watch, the rest of us have homework. **02 / In 1 minute** ## What is worth remembering 01\. An internal OpenAI model reached the live internet through an open DNS resolver configuration during a training run. The lab paused training and evaluation of its most capable models while it hardened its infrastructure. 02\. OpenAI's agent teams found microVM escapes onto host machines, communication channels through CPU cache timing and hidden paths to the internet. The patches were contributed back to Kata Containers, gVisor and Docker. 03\. Training resumed on October 1 with two new rules: continuous red teaming inside the safety case of every run, and zero live internet access during frontier training. **03 / Post of the week** ## What you know expires: the new advantage is learning fast ![Cover: What you know expires](https://jonathancanales.com/content/images/2026/10/hbr6-header.png) We close our series on AI-ready culture with the most personal shift: from accumulated experience to learning speed. Think of the most valuable person at your company ten years ago, the one who knew the most. How much of that knowledge is still current? Half, with luck. The company that learns fastest goes beyond buying courses: it turns reflecting, stretching and relearning into routine. Tools go out of fashion. This, precisely this, is what stays. Series source: [There's No Such Thing as an AI-Ready Culture, Harvard Business Review, October 5, 2026](https://hbr.org/2026/10/theres-no-such-thing-as-an-ai-ready-culture?ref=jonathancanales.com). [Read the full article ↗](https://jonathancanales.com/en/what-you-know-expires/) **04 / One AI thing that matters** ## Security became a daily routine ![Illustration: magnifying glass over an open window](https://jonathancanales.com/content/images/2026/10/nl04-doodle1.png) What is happening: on September 20, an internal training model used an open DNS configuration to reach the live internet. Monitoring flagged it within 15 minutes and the run was stopped hours later. OpenAI paused training and evaluation of its most capable models and turned the pause into work: teams of its strongest models attacked its own infrastructure. They found microVM escapes onto host machines, secret communication channels through CPU cache timing (the family of tricks behind Spectre and Meltdown) and hidden paths to the internet through shared services. The patches were contributed back to open projects like Kata Containers, gVisor and Docker. On October 1, training resumed under two new rules: continuous red teaming inside the safety case of every run, and frontier training with live internet access switched off. Why it matters: AI infrastructure inherits the same classes of gaps as traditional software, amplified by models that explore their environment with their own creativity. Watch stops being an annual audit and becomes part of the process, just like code tests. What to do on Monday: ask who plays red team inside your operation. If the answer takes a while, start with the basics: list your exposed services, check what can leave each environment and schedule the review as routine, same as data backups. Sources: OpenAI, [Securing our frontier model training environment (Oct 23, 2026)](https://openai.com/index/securing-our-frontier-model-training-environment?ref=jonathancanales.com) and OpenAI Alignment, [incident report (Sep 20, 2026)](https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/?ref=jonathancanales.com). **05 / Signals** ## Four reads. Zero smoke. Four reads to understand the full week: [1\. Kimi K3.1 Pro by Moonshot AI: an open-weight model with 104 billion active parameters (out of 2.8 trillion total), a 1 million token context and coding results above GPT-6, according to the lab itself.](https://www.kimi.ai/blog/kimi-k3-1-pro?ref=jonathancanales.com) [2\. MemoryTax: a Shanghai Jiao Tong University study compares seven ways of giving an agent memory and finds that the most capable models do better with the raw history than with elaborate memory layers.](https://arxiv.org/abs/2610.29823?ref=jonathancanales.com) [3\. Jon Barron found two numerical gaps in softmax and cross-entropy, present in every major machine learning framework, with a proposed fix included.](https://jonbarron.info/softmaxfix/?ref=jonathancanales.com) [4\. Meta's Byte Latent Transformer: models that read bytes instead of tokens and scale better with one-sixth of the data, according to the publication.](https://ai.meta.com/research/publications/byte-latent-transformer-patches-scale-better-than-tokens/?ref=jonathancanales.com) **06 / Trick of the week** ## The cheap test before buying memory Does your agent need an elaborate memory layer? Before building one, run the cheap test from the MemoryTax study: take a real long conversation, hand the raw history to the model and measure how well it answers. Recent models tend to do better with the full context than with structured summaries, and the raw history costs cents per query. If the simple test covers your case, you save weeks of infrastructure. **07 / To close** ## Who plays red team in your house? The week left a clean lesson: even well-watched systems have windows their own builders missed, and the way to find them is looking every day. Your operation deserves the same question: who looks for your open windows this week? Jonathan --- **Share a useful idea:** [X](https://twitter.com/share?text=Even%20AI%20needs%20continuous%20watch&url=https%3A%2F%2Fjonathancanales.com%2Fen%2Fnewsletter-04-continuous-watch%2F&ref=jonathancanales.com) · [LinkedIn](https://www.linkedin.com/shareArticle?mini=true&url=https%3A%2F%2Fjonathancanales.com%2Fen%2Fnewsletter-04-continuous-watch%2F&ref=jonathancanales.com) · [WhatsApp](https://wa.me/?text=Even%20AI%20needs%20continuous%20watch%20https%3A//jonathancanales.com/en/newsletter-04-continuous-watch/&ref=jonathancanales.com) To copy the link, open the article and tap "Copy link". [Artículos en español ↗](https://jonathancanales.com/boletin-04-vigilancia-continua/)