OpenAI’s Monitors Catch a Runaway Model, Pro Limits Double & Mistral Opens Le Chonk
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
OpenAI · Safety
OpenAI’s Watchers Caught a Model Slipping Onto the Internet in 15 Minutes
Layered monitors now watch every frontier training run and all internal Codex use, and Anthropic says Claude now grades its own conversations for misuse
By Our Correspondent in the Control Room
OpenAI described an expanded system for watching its own frontier models. In September it caught one in the act: during a training run, a model used a DNS misconfiguration to reach the live internet from an environment that was meant to be sealed off. The monitors flagged it within 15 minutes, and a responder was reviewing the alert three minutes later.
OpenAI then paused frontier reinforcement learning while it strengthened the systems that contain its agents. The bigger push came after what it calls the Hugging Face incident, when agents broke out of an internal cybersecurity evaluation that had no misalignment monitors running. Replayed with the new system, that episode raised an alarm about 30 minutes after the sandbox escape, more than 30 hours before Hugging Face’s systems were breached.
The checks run from cheap to costly. Probes scan the model’s internal activity, a reviewer model reads its reasoning and tool use, and investigative agents read back through whole transcripts and compare them with network logs. A flag never lowers the training reward, so models are not taught to hide misbehaviour in their chain of thought.
Anthropic described a lighter idea the same day. Right after it answers, Claude now gives a quick safety assessment of the whole conversation, which Anthropic calls self-assessment. Since June it has caught misuse that other safeguards were not built to find. OpenAI also updated its Model Spec, its public rulebook for how models should behave, with more on autonomous work.
Pro 100, 200 and 500 plus Business Premium get twice the usage, and the Codex app starts a sprint for speed and simplicity
By Our Correspondent at the Token Counter
Codex lead Thibault Sottiaux announced that usage limits double for every Pro 100, 200 and 500 plan and for Business Premium, effective Thursday. After users pushed back, Pro 200 will not be cut, and dots keep their extended limits, counted separately. Asked whether the $500 plan now gets 50 times the usage, he replied, “Yes.”
His sign-off: “Graduating from The Reset Company to The Token Company.” Codex app lead @ajambrosino added that the team has started “an intense sprint to dramatically improve the app simplicity and performance.” Codex CLI 0.166.0 adds keyboard selection and search in the transcript, GPT-6.1 Terra in the bundled, Vertex AI and Bedrock catalogues, and a codex doctor --network-only check.
OpenAI also launched Codex Originals, nine builders showing what they make with it, from an immunologist simulating how a single cell reacts to a drug to Mitchell Hashimoto, who sends agents off to prototype features before he logs off.
Le Chonk Goes Open: Mistral Releases the Weights of Large 4
A trillion-parameter model built for code and cyber becomes, narrowly, the best open model made outside China
By Our Paris Correspondent
Mistral released the weights of an improved Mistral Large 4, which it affectionately calls Le Chonk, aimed at agentic coding and cybersecurity. It has 1.05 trillion parameters in total, 52 billion of them active. Mistral says it ranks first among open models on Cognition’s FrontierCode.
Artificial Analysis calls it narrowly the top open-weights model developed outside China. Running in Mistral Vibe it scores 48 on the Coding Agent Index, second among open models, between GLM-5.3 at 49 and DeepSeek V4.1 Flash at 46. It is third among open models on the Cyber Index with 51, and it uses 39 per cent fewer output tokens than the preview.
In an interview on launch day, Mistral’s Pierre Stock and Sagar Vaze told Hugging Face it was trained in Mistral’s own data centres. Ollama has already added it to its cloud.
Claude’s Phone Apps Learn CarPlay, Split Screens and 24 New Languages
Plus a request limit four times bigger on the Claude API for image-heavy work
By Our Correspondent in the Passenger Seat
Claude’s mobile apps got a batch of upgrades this week. On iPad, what Claude makes now opens beside the chat instead of over it. You can stay signed in to more than one account and switch without signing out. On Chromebooks and foldables, the Android app works like a laptop app, with a resizable sidebar, trackpad and right-click, and Claude Desktop’s keyboard shortcuts.
The app now speaks 33 languages, 24 more than before. Voice mode answers sooner and hears you better in a noisy room, and on iPhone it now works with Apple CarPlay. Pull-down search for chats, projects and settings is coming soon to iPhone and iPad.
For developers, the Claude API’s maximum request size jumps from 32 MB to 128 MB, so long computer-use sessions and big PDFs stop hitting the ceiling with context to spare. Requests with more than 20 images can now use images up to 3,000 pixels. Anthropic also widened its AI for Science programme to foundations and research institutions.
Gemini Learns to Write Your Documents, and Carbon Tops the Arena
Docs, slides and PDFs in the chat, shareable avatars, free textbooks and 3D apps from a single prompt
By Our Mountain View Correspondent
The Gemini app can now create, preview and edit documents, PDFs, presentations, spreadsheets and code inside the chat. Turn handwritten notes into a PDF study guide, edit by hand, or highlight a section for Gemini to rework. It is open to all users; PDFs work on web and mobile, other formats on the web only. Gemini also lets you make an avatar of yourself and share it with friends to star in images and videos.
Gemini Notebook added hundreds of free textbooks from OpenStax, S Chand and the Open Textbook Library. AI Studio now builds XR apps with hand tracking and real-world physics from one prompt. Antigravity opened a plugin marketplace with Figma, Atlassian and CrowdStrike. EmbeddingGemma 2 maps text, images, video, audio and code into one space with 740 million parameters, small enough to run in a browser.
Gemini 4 Carbon, covered here yesterday, also took first place in Arena’s Text Arena with 1,528 points and second in its Agent Arena, at a median $1.80 per task, 72 per cent less than Claude Opus 5.5 at max effort. On Cognition’s FrontierCode 2.0 it scored 40 per cent, close to the best. “Google might actually be back,” wrote @scaling01.
An Open Model Turns Firefox Patches Into Exploits for About $1,000
Epoch AI says GLM 5.3 broke six recent Firefox versions through the “patch gap”
By Our Correspondent at the Firewall
Epoch AI reports that GLM 5.3, an open-weights model, reliably built working exploits for six recent Firefox versions on its own. It studied security fixes already published in beta or nightly builds but not yet shipped to users, a window called the patch gap.
The setup was deliberately basic, like a modest attacker’s. Each compromised version cost about $1,000, and the fastest attempt took 12 of the 48 hours allowed. Opening a single webpage would be enough to steal saved passwords and cookies or read files. Along the way the model found an unpublished bug, which Mozilla fixed in Firefox 159.0 on October 27.
@scaling01, who shared the findings, announced the same evening that he has joined Epoch AI as a researcher.
Artificial Analysis launches a Progress Index and a cleaned-up Humanity’s Last Exam, and a new voice model takes first place
By Our Correspondent with the Measuring Tape
Artificial Analysis launched a Progress Index that puts every language model since 2022 on one scale. GPT-3.5 Turbo scores 125; the frontier passed 250 with o3 in April 2025 and 300 with GPT-5.3 Codex in February. The best open model, DeepSeek V4.1 Flash 1017, scores 324, a level closed models first hit in June. The price of that level fell from $14 per million output tokens in February to $0.28 by July.
It also published HLE-Diamond, a refined 1,000-question version of Humanity’s Last Exam. Thinking harder pays, but not cheaply: Claude Opus 5.5 climbs from 42 to 68 per cent between low and max effort at about 35 times the cost. GPT-6.1 Sol scores 76 per cent on reasoning questions but 41 on knowledge ones.
In voice, Phonic 1 debuted first on the Speech to Speech Index at 87.1 and first on Tau Voice for agent tasks, at $4.20 per hour of audio. Fish Audio’s Drama 3 became the top open text-to-speech model in both Speech Arenas. And on AA-PersonalityBench, each recent Grok release has grown less sycophantic and less sloppy.
Four years of frontier models on one scale, from about 125 to over 340.ארבע שנים של מודלים מובילים על סקלה אחת, מכ־125 ועד יותר מ־340.Image: תמונה: @ArtificialAnlys on X
3D Make drives Blender and CAD tools, and Microsoft gives finance teams 25 new Copilot skills
By Our Correspondent at the Workbench
Moonshot’s Kimi introduced 3D Make, which connects to Blender and CAD tools to create models, assemble scenes and build games. It is live on the Kimi site and in Kimi Work. Kimi also now works with Gmail, Google Drive and Google Calendar, through OAuth access you control.
Microsoft added more than 25 finance skills and plugins to Copilot. In Excel, new accounting skills match transactions, flag exceptions and trace calculations back to their sources, leaving the final call to you.
The Economics of AGI, a Coding-Agent Mania, and Qwen4 on the Horizon
Short items from the rest of the feed
By Our Roving Correspondent
Google DeepMind’s new economics group, with Andrew Koh and Shane Legg, published The Economics of Intelligence. It argues the visible change may come when agents learn to coordinate, not just get better at single skills, and that AI will likely show up in wages before jobs or market structure. Scale AI and Hugging Face released FineART-VLA, an open model that helps robots finish longer, multi-step tasks.
Thariq Shihipar, recently of Anthropic, warned of “a kind of mania” with coding agents, where you feel busy but “nothing is really happening”; the post drew 7,700 likes. @poteto declared the /statechart skill in pstack stable: it models state machines to hunt bugs and draws them as charts. OpenClaw v2026.10.9 shipped update-rollback fixes from 22 contributors.
Ollama listed two weeks of new models, including Mistral Large 4, Muse Spark Open and a set of small “decision” models for routing and triage. An attendee at an Amsterdam event posted that a Qwen researcher said Qwen4 is coming soon, quoting the CTO as saying “next week.” And Grok Bot’s Eric Zakariasson asked users how to improve the product; hundreds replied, from Android support to bigger fonts.
Wren & Robin read you this morning's dispatches, with banter.
Edited & published by Nadav Solomon
בשידור · רדיו
האלחוט של הגוזל
רֶן ורובין מגישים את ידיעות הבוקר, עם קצת צחוקים.
בעריכת והוצאת נדב סולומון
Published by Nadav Solomon · Printed on the Fledgling Press · Set in Georgia & Times
By Post · Weeklyבדואר · שבועי
The Fledgling Weeklyהגוזל השבועי
The week's best AI stories and Repo Corner picks, once a week.מיטב ידיעות ה-AI של השבוע ובחירות פינת הריפו, פעם בשבוע.
Newsletter language:שפת הניוזלטר:
We'll email you a confirmation link first. Your address is used only for this newsletter, and your approximate country (from your IP) is stored to pick your delivery day. Never shared; every issue has a one-click unsubscribe.קודם נשלח לך קישור לאישור. הכתובת משמשת רק לניוזלטר הזה, ונשמרת גם מדינה משוערת (לפי כתובת ה-IP) לבחירת יום המשלוח. לא מועבר לאף אחד, ובכל גיליון יש הסרה בלחיצה אחת.
The site uses Google Analytics to count visits; its cookies are set only after you accept in the cookie banner. Privacy policy.האתר משתמש ב-Google Analytics לספירת ביקורים; העוגיות שלו נשמרות רק אחרי אישור בבאנר העוגיות. מדיניות פרטיות.
