Three stories this week share a shape, even though none of the people involved were talking to each other. Mistral handed out the keys to its biggest model, built to be just as good at defending a network as at writing code for one. Independent security researchers found a way for one ChatGPT user's private conversation to leak into a stranger's browser, through a feature that was allowed to touch both untrusted content and private data at once. And an Indian company signed on to build AI for the Army, Navy and Air Force on infrastructure that never leaves the country. None of this is our work. All of it is the same argument we keep making, arriving this week from three directions that had no reason to agree with each other.
Why we think this matters
The argument is simple and it keeps getting proven the hard way: a system that can run entirely inside your own walls is no longer a lesser system, and a system that is allowed to do too much at once is where the trouble starts.
For years, the trade-off was real. Keep a model inside your own building, off the public internet, and you gave something up — speed, polish, capability. That trade-off is closing fast. This week's most capable newly open model is also one of the most capable at security work, available to download and run rather than rent. The gap between "sovereign" and "state of the art" keeps narrowing, which is exactly the bet this company is built on.
The model that can help a security team is the same model that can help an attacker. That is not a reason to avoid capable AI. It is the reason every system we build has a boundary around what it is allowed to do on its own, and a record of what it actually did.
The ChatGPT story makes the other half of the argument. The bug was not that the AI was too capable — it was that one feature was allowed to handle an untrusted web page and a person's private account in the same breath, and that was enough for one to reach the other. That is the exact failure we design against: no part of a system touches something private and something unverified at the same time without a person standing in the gap. It is reassuring, if not exactly cheering, to watch the rest of the industry relearn why that boundary matters.
What matters this week
- Mistral released the full weights of its largest model, tuned for both coding and defensive cybersecurity work, so anyone can run it on their own machines instead of renting access to it. Confirmed by Mistral's own announcement, 29 October 2026 (mistral.ai). Mistral's claims about how good it is at security work are the maker's own and have not been independently checked yet; two technology publications and ENISA, the EU's cybersecurity agency, have flagged that a model this capable at defensive security could also be misused offensively, and ENISA says it is now testing the model itself, with no results published yet. What it means for us: proof that "runs entirely on hardware you own" and "genuinely capable" are no longer opposites — and a reminder that capability alone is not the whole story; what a system is allowed to do with it is.
- Independent security researchers found a chain of bugs in ChatGPT that could let an attacker take over another person's account and read private conversations, just by getting them to open a crafted web page. OpenAI has fixed it and paid the researchers for finding it. Confirmed, reported by the researchers themselves and relayed by independent commentator Simon Willison, 31 October 2026 (simonwillison.net). What it means for us: a textbook case for drawing a hard line between a system that reads untrusted content and a system that holds something private, rather than trusting one feature to handle both safely.
- OpenAI released a new, cheaper version of the ChatGPT model it recommends for letting an AI operate a computer directly — clicking, typing and filling in forms — instead of only answering in text. Confirmed by OpenAI's own announcement, 26 October 2026 (openai.com). OpenAI's claims about how much faster and more capable it is are the maker's own; an independent group that tested it themselves found it a solid, unremarkable performer rather than the clear leader OpenAI's own figures suggest, and OpenAI's own safety paperwork separately rates its hacking capability at the company's highest internal concern level. What it means for us: nothing changes for us this week, but an AI that can drive ordinary business software on someone's behalf keeps getting cheaper, which is worth tracking as a building block.
- Sarvam, an Indian AI company, signed a five-year agreement with the headquarters that coordinates India's Army, Navy and Air Force, to build AI for training, decision support and related work across all three services — running on infrastructure built and kept in India. Confirmed by Sarvam's own announcement, 27 October 2026 (sarvam.ai), which states that people keep final say over operational decisions. What it means for us: not a deal that touches us, but a real, signed example of the case we make every day — that an Indian organisation can build and keep AI on its own ground rather than renting it from somewhere else.
New models and price moves this week
Moonshot AI, the Chinese lab behind the Kimi models, released Kimi K3.1 Pro, a new flagship aimed at coding, design work and research, on 23 October 2026, and says it will publish the model's full weights — letting anyone download and run it themselves — within the next few days. Moonshot says it performs close to Anthropic's Claude Opus 5.5 on its own tests, at a noticeably lower price; nobody outside the company has checked that comparison yet. Separately, OpenAI said it will start matching the pay-as-you-go credit many developers use for its highest-paying ChatGPT subscribers, confirmed by OpenAI's own documentation, 29 October 2026 — the same kind of offer Anthropic already makes, and one more sign the large labs are competing hard on price for anyone building directly on their platforms. And Anthropic is giving government agencies a meaningfully lower price on Claude for enterprise and government use, confirmed by Anthropic's own announcement, 27 October 2026 — a US government story mostly, and nothing that changes anything for us today.
What people are trying and arguing about this week
GitHub's own 2026 look-back at its platform, published 28 October 2026, reports that AI agents now start a large and fast-growing share of all proposed code changes on the service, up sharply from where that share stood at the start of the year, and that the total amount of code merged on the platform has roughly tripled in two years. That is GitHub measuring its own platform, not an outside audit, and a trade publication covering the report cautions that it shows how much code agents are producing and how quickly it gets merged — not whether that code is actually good; nobody has published an independent check of that yet. Worth watching, not yet worth trusting blindly: it says agentic coding has gone mainstream on real projects, not that it is ready to skip review. Separately, a startup called Liquid AI released a new family of very small AI models, small enough to run entirely on an ordinary laptop or office computer rather than a data centre, aimed at everyday jobs like sorting an inbox or managing a calendar. Liquid's own claims about speed and efficiency are the maker's own; nobody outside the company has tested them independently yet. It is another sign that "small enough to run on hardware you already own" keeps getting more capable — the same bet behind VarahiEdge — though there is nothing here to use today.
Explained simply: JEPA and world models
In one line. JEPA is a different way of training an AI model: instead of learning to guess the exact next word or pixel, it learns to guess the gist of what comes next.
An everyday picture. Today's chatbots learn the way a very good autocomplete does — filling in the next word, over and over, millions of times, until the exact wording becomes predictable. A JEPA model learns more like recognising that two photographs are of the same street corner taken from different angles, even though none of the actual pixels line up. It is being taught to recognise sameness of meaning, not sameness of wording.
Why people care. The idea's loudest backer is Yann LeCun, the scientist who led Meta's AI research until he left to pursue this approach full time at his own company, which has raised well over a billion dollars doing it. His argument: training a model to predict the exact next word wastes effort memorising details that do not matter, while training it to predict meaning should need far less data and hold up better in situations it has never seen.
What is real and what is hype. This week, LeCun and a co-author published a paper offering something closer to mathematical proof, not just intuition, that two JEPA models trained separately on different kinds of data — one on images, one on captions — end up understanding the world the same way, provided each is shown genuinely different views of the same thing. They backed it with their own tests across a very large set of real photographs, where a JEPA model far smaller than today's big chatbots matched or beat them at linking a sentence to the photo it describes. That is a real, published result — but it comes from the method's own co-inventor, has not yet been reproduced by an outside lab, and a separate paper at a major AI conference this year argues that the apparent agreement between models mostly disappears once it is measured more carefully. Nothing built on this idea is for sale yet.
What it means for us. Watch, don't build. If the idea holds up outside its inventor's own papers, models that need far less data to understand their surroundings would matter a great deal for AI that has to run on a small, offline computer in a factory — which is exactly the situation our own engineering AI is built for. That is a reason to keep reading, not a reason to change anything we ship this week.
Nothing here needs action this week. If one thing is worth carrying into a customer conversation, it is the Mistral release: a fully self-hosted model that is also genuinely good at defensive security work is a new combination, worth mentioning to anyone who still assumes "runs entirely inside your own building" means "give something up." We will keep reading the same sources next week, and if nothing moved, we will say so instead of padding it.
Read about AI Transformation →