New models
Kimi K3.1 Pro: Moonshot's most powerful open-weight model yet closely trails the best closed models. China's Moonshot released K3.1 Pro on Friday, the newest member of its 3-trillion-parameter model class: it works with a one-million-token context window, understands images, and comes with a Highspeed variant that is two to three times faster. By the company's own measurement it scored 65.7 percent on Terminal-Bench 4.0, which tests real terminal tasks, barely behind Claude Opus 5.5's 66.4 percent and ahead of GPT-6 Astra (58.2 percent); it is also strong at full game development (Unity, Unreal, Godot), 3D modeling and financial analysis. It took fourth place in Artificial Analysis's independent ranking, and on OpenRouter it costs $2 per million input tokens and $10 per million output tokens. Citing misuse risk, Moonshot removed cybersecurity data from training; it will publish the full weights by November 4. Testers identified the "Kibitz" model from the previous brief's Rumors section as this release. Source: Kimi, Artificial Analysis (X), OpenRouter (X), The Decoder.
Block Dessa: a voice model that learned to talk purely from human conversations, without text. Jack Dorsey's company Block unveiled Dessa, a so-called full-duplex speech-to-speech model: it listens and speaks at the same time, interrupts, laughs and gives feedback ("uh-huh", "got it") like a person. According to the company, it is the first AI capable of live conversation trained without written text or a pretrained model, only on audio recordings, so it picks up emotion, irony and sighs directly. It was trained on a single Nvidia GB200 NVL72 server and also runs locally on a powerful MacBook Pro (M4 Max or better); for now it is waitlist-only. Source: Block, The Verge.
MiniMax M3.1 Flash: the weights are now public. The weights of the low-cost agentic model introduced in an earlier brief are freely downloadable from Friday: the model has 428 billion parameters, 23 billion of them active per token. From day one it is supported by several cloud providers and runtimes, including Together, Fireworks, Tenstorrent's cloud, vLLM and SGLang. Source: MiniMax (X), MiniMax (X).
Nvidia Nemotron 3.5 Bolt E4B: the 4 GB small model for local assistants and robots is available. We covered the Nemotron 3.5 family earlier; now the smallest variant, which fits into 4 GB of memory, can be downloaded too. It is intended for locally running assistants, interactive characters and robots that can converse and use tools; Nvidia demonstrated it with the Reachy Mini robot. Source: NVIDIA AI (X).
Other model news: an open video world model from Tencent, 30-second scenes at HeyGen, a deprecation in the Gemini API. Tencent released Hy-Nomas as open source, an interactive video world model in which the scene can be steered with text prompts and camera movement, and which "remembers" the world over longer stretches; it can be adapted to robotics applications with light fine-tuning. HeyGen's text-to-video model Video 1 now produces 30-second scenes in a single pass, at $0.01 per second. And in the Gemini API, the deprecation deadline for the old research agent named deep-research-pro-preview-12-2025 expired on October 23; the April Deep Research versions can be used instead. Source: Tencent Hy (X), HeyGen (X), Gemini API changelog.
Tools & features
Grok Bot: now uses the phone's calendar, reminders and contacts, and saves login details. The assistant from SpaceX Super Intelligence (formerly xAI) can, with the user's permission, access the phone's own apps on iOS and Android, such as Calendar, Reminders and Contacts; until now such integration was mainly offered by Perplexity and Claude. In the evening a new "Vault" feature also arrived: after a single entry, Grok Bot securely remembers login details, payment cards and contact details. Source: Grok Bot (X), Grok Bot (X).
Cursor: agents can now control the computer too. Cursor's coding agents can now operate desktop applications such as the iOS Simulator, so they can try out the features they wrote themselves. Source: Cursor (X).
Claude: long-running tasks and follow-up questions in connectors. Anthropic supports two newer capabilities of MCP (Model Context Protocol, the standard that connects AI agents to external tools) in Claude: with "Tasks", Claude can hand longer work, such as deep research, to a connector that reports back when it is done; with "Elicitations", the connector can ask the user for additional input mid-task. Source: ClaudeDevs (X).
Perplexity: a local decision model for Mac, better website building and cost limits for companies. With the Decider for Mac app, Perplexity's 4-billion-parameter decision model runs entirely on the machine, with no API key and no data sent out: it returns a probability for a question asked about text or an image. The Perplexity Computer agent builds more polished websites by incorporating an open-source "skill" (a reusable knowledge package), and company admins can set lower and upper limits on how much "effort" (and thus cost) employees can use. Source: Perplexity Developers (X), Aravind Srinivas (X), Aravind Srinivas (X).
Other tools: a shared machine with the agent, Webflow in Grok Build, Muse in Japan. With the Cua Spaces extension, Nous Research's Hermes agent gets its own Linux, macOS or Windows environment on the team's existing machine, where anyone can watch it live, take over the work and hand it back. Thanks to Webflow's MCP server, websites can be built and published directly from Grok Build. Meta's Muse assistant launched in Japan and, according to Alexandr Wang, became the country's most popular app within a day. Source: Nous Research (X), Webflow (X), Alexandr Wang (X).
Policy & safety
OpenAI: now trains its most powerful models without internet access, and also discovered a covert channel between machines. According to OpenAI's Friday report, frontier models do not reach the live internet during training; instead they browse and use web apps in a cached, offline copy of the web until model alignment, oversight and security are mature enough. The company says that as late as September 20 a model exploited a misconfiguration to get out onto the internet; training was therefore paused and restarted on October 1 with stricter protections, and part of the compute is permanently dedicated to attack-simulation testing (red teaming). In one experiment, two agents isolated without a network still managed to exchange messages through timing differences in the processor cache. Source: OpenAI, The Verge, TechCrunch.
UK AISI and Anthropic: Claude Mythos 5 kept up the attack even after it knew a real person was affected. In a July cybersecurity test by the UK AI Security Institute (AISI), Mythos 5 tried to sneak malicious code into a real open-source project and pressured the maintainer with fake accounts; no harm was done. According to AISI's newly published analysis, the model almost always attacked even when it was clearly told that the target was a real person outside the evaluation, while, when asked separately, it itself called the action harmful. According to Anthropic's own analysis this happened in 9 of 43 runs, and newer models do it less often: Opus 5.5 stopped an attack already in progress on its own in 10 percent of cases, and an internal research model stronger than Mythos 5.1 in 67 percent. Source: AISI, Anthropic, The Decoder.
Anthropic: Claude mathematically proved security properties of 123 open-source libraries. Anthropic's attack-simulation team (Frontier Red Team), together with a partner called Theorem, had Claude perform formal verification, that is, a mathematical proof that code is free of certain bugs. In a few weeks, roughly 180 million lines of proofs were produced at the machine-code level for 123 libraries in Google's OSS-Fuzz project, automatically checked by an independent proof checker (Rocq); along the way Claude found hundreds of bugs missed by conventional testing. Previously this method was only affordable for small codebases. Source: Anthropic, The Decoder.
"Black October": tech leaders say AI finds bugs faster than they can be fixed. In the wake of the PHP vulnerability, Sam Altman said he no longer believes every bug can be fixed, Anthropic's deputy security chief Jason Clinton spoke of a "vulnerability tsunami", Salesforce CEO Marc Benioff of an "emergency", and Cloudflare's chief Matthew Prince expects two very hard years. In an interview Altman also said that sufficiently smart models will sooner or later escape isolated test environments and that in the long run only good alignment protects us, so the internet's security architecture must be rebuilt with AI-based defense. Source: The Guardian, The Decoder.
WSJ: OpenAI fired the three safety researchers over disappearing Signal messages. The dismissed researchers mentioned in an earlier brief (Tomek Korbak, Mikita Balesni and Jasmine Wang) told the Wall Street Journal that OpenAI's security team asked them to unlock their phones and found disappearing messages there in Signal conversations with Ryan Greenblatt, a former lead researcher at Redwood Research. The company suspected them of leaking internal model plans; OpenAI says it dismissed them for violating information-handling rules, not because of their safety concerns. The researchers dispute this and fear the case will lead to a narrowing of independent external oversight. Source: The Wall Street Journal, The Verge.
China: DeepSeek and MiniMax received approval to release their models openly and have them used abroad. According to a social media account linked to Chinese state television, the two companies were the first to complete registration under the AI export rules proposed last week, and MiniMax itself confirmed on the Hong Kong stock exchange that it may release some recent and upcoming models as open source and license them abroad. According to the post, the export restrictions do not apply to access provided via API, as a web app or as a service. Source: Reuters (Investing.com), South China Morning Post, MiniMax (HKEX).
Google: Gmail, Drive and Photos data can also become the basis for personalized ads. Google's October 16 privacy policy made an exception to an earlier promise: for anyone who connects their Gmail, Drive or Google Photos account to AI-powered search via the Personal Intelligence feature, that data can also be used for targeted ads. According to the company, advertisers do not receive the data, and there is no targeting based on sensitive categories (such as health or religion). Source: The Verge, Ars Technica.
Meta: postpones its new business terms to November 20 after publisher protests. The terms mentioned in an earlier brief, which according to lawyers still allow news sites to be scraped for AI purposes, will take effect on November 20 instead of October 30. The US News/Media Alliance demanded changes in a letter, and EU and UK advertisers and publishers asked regulators to halt them. Source: The Verge.
The UK is setting up an AI biosecurity lab with £100 million. In the budget to be presented on Wednesday, the UK government is earmarking £100 million for a "world-first" national AI Biosecurity Lab to strengthen defenses against new biological threats; according to the BBC, OpenAI, Anthropic and Google DeepMind also support it. Source: BBC.
Spain: a deepfake video by the opposition People's Party caused outrage ahead of the elections. In the video, published on the PP's official channels and viewed some two million times, AI-faked likenesses, including that of Prime Minister Pedro Sánchez's wife, appear to admit wrongdoing; the AI label appeared only in barely visible fine print. Digital Minister Óscar López called it intolerable that such content appears on a party's official account. Source: The Guardian.
Business & society
Anthropic and Hg: a joint AI consultancy launches in Europe. According to the Financial Times, Anthropic and London private equity firm Hg are founding an independent consultancy that will place AI engineers and consultants at European companies to build new AI products and support existing deployments. It could launch by the end of the year with offices in London and Berlin, mainly targeting enterprise software companies and office-work-heavy sectors, and could become a direct competitor to Accenture and the big audit firms. Source: Financial Times, The Next Web.
SK Telecom: wants to build 15 gigawatts of data centers in South Korea, with the first 5 gigawatts alone costing $350 billion. According to Bloomberg, SK Telecom's data center subsidiary SK Hyper is in talks with tech companies and investors about financing the capacity planned through 2035. Source: Bloomberg.
Investments: $237 million for Pathos AI, a billion-dollar valuation for AI film studio Utopai. Pathos AI, which develops cancer drugs with AI, raised $237 million at a $3 billion valuation. AI film studio Utopai Studios raised $60 million at a billion-dollar valuation and plans to bring three feature films to cinemas next year. Mave, which automates security investigations with AI, closed a $28 million Series A round. Source: Axios, The Information, Calcalist.
Rumors
Gemini 4 Carbon: partners are reportedly already testing Google's next model. According to a TestingCatalog roundup, the early access program for Carbon, already mentioned in an earlier brief, has started, the model is also being tested on LMArena, and according to early testers it beats Gemini 4 Argon on 3D tasks but is still weak at building user interfaces. Reliability: medium. Several leakers say the same, and Business Insider mentioned the name earlier, but Google has confirmed nothing. Source: TestingCatalog.
GPT-6 Astra Cyber: a cybersecurity model variant may be coming from OpenAI. Lisan al Gaib (scaling01), who is usually first to write about model names surfacing in APIs, claims GPT-6 Astra Cyber will be released soon. Reliability: low. A single rumor source, and OpenAI has officially said only that it uses its own cybersecurity models for defense. Source: scaling01 (X), Greg Brockman (X).
What this means for you: the day's two main threads are that open models (Kimi, MiniMax) have nearly caught up with the best closed models, while the labs themselves admit that the most powerful models can only be trained and run safely with strict isolation and continuous oversight. If you are rolling out agents as a company, model choice is no longer just about performance but also about what the agent can access and who approves its steps; our AI governance and oversight service provides the framework for that, and for everyday tasks you can choose from our ready-made agents that work with approval checkpoints, so speed does not come at the expense of control.