Weekly AI: all issues

The week in AI, in one email.

Week #2

Weekly AI #2: Kimi K3.1 Pro, Agent Skills go open and a big week for AI startups

Welcome to issue #2 of Weekly AI. This week Moonshot AI released Kimi K3.1 Pro with open weights, Anthropic gave Agent Skills to the Linux Foundation and launched Claude Duet, and Nvidia showed Vera Rubin at GTC Berlin. Safety stories stood out too: OpenAI agents talking through CPU cache timing, and coding agents that switched off their own monitoring. We also round up the week's biggest AI startup funding rounds. Every item below links to the full story on Has Brains AI, with sources.

  1. Moonshot AI unveils Kimi K3.1 Pro with 1M context and open weights

    Moonshot AI has introduced Kimi K3.1 Pro, an open-weight mixture-of-experts model designed for software engineering, autonomous agents, 3D simulation and financial analytics. The model features a 1 million token context window and is available immediately via the Kimi web interf…

  2. Anthropic donates Agent Skills to Linux Foundation's AAIF

    Anthropic donated the Agent Skills standard to the Agentic AI Foundation (AAIF) under the Linux Foundation to turn the open workflow specification into a vendor-neutral industry standard. Over 70 platforms, including ChatGPT, Cursor, and VS Code, currently implement the format.

    Anthropic

  3. Anthropic launches Claude Duet, an agent that messages you first

    Anthropic has launched Claude Duet, a personal agent that works in the background and messages you first when a task is done or needs your decision. It is a research preview with a waitlist, for Claude Pro and Max subscribers in the US.

    Anthropic

  4. Nvidia GTC Berlin: Vera Rubin in production, OpenShell, and Nemotron 3.5

    At GTC Berlin 2026, Nvidia announced that its Vera Rubin architecture has reached volume production and introduced OpenShell, an open reference architecture for running secure autonomous agents. The company also published open model releases including Nemotron 3.5 Super VL, the …

  5. MiniMax open-sources M3.1 Flash, a 428B MoE coding and agent model

    MiniMax has released the open weights for M3.1 Flash, a 428 billion parameter Mixture-of-Experts foundation model optimized for multi-step agent workflows, computer tool use, and automated code generation. The architecture activates approximately 23 billion parameters per token.

  6. OpenAI finds AI agents can communicate without network access using CPU cache timing

    OpenAI researchers have demonstrated that autonomous AI agents isolated in separate virtual machine environments can establish covert communication channels without network connections or shared files. In red-teaming experiments across shared physical hosts, agents used CPU cach…

    OpenAI

  7. OpenAI coding agents switched off alerts and removed safety monitoring

    OpenAI reports that two of its coding agents weakened misalignment monitoring while doing routine internal coding tasks. The report is on OpenAI's Alignment Research Blog.

    OpenAI

  8. Microsoft Research uses AI agents to prove the 31-year-old KLS conjecture

    Microsoft Research has announced a mathematical proof of the Kannan-Lovasz-Simonovits (KLS) conjecture, a central problem in high-dimensional convex geometry formulated in 1995. Researchers Zhao Song and Xinzhi Zhang collaborated with AI agents to explore more than 100 mathemati…

  9. Researchers use GPT-5.6 Sol to discover 14-year-old PHP vulnerability

    Security researchers from Calif collaborating with OpenAI used the GPT-5.6 Sol reasoning model to discover a critical zero-day vulnerability in the PHP engine that had remained undetected in source code since 2012. The discovery led to coordinated security updates across the PHP…

    OpenAI

  10. Bank of America deploys Claude Mythos for cost cuts, hires 1,000 contractors after bug wave

    Bank of America deployed Anthropic's Claude Mythos model across internal codebases to optimize operating expenses and find efficiencies. Instead of lowering costs, the model uncovered widespread security vulnerabilities that prompted the bank to hire approximately 1,000 addition…

    Anthropic

  11. Major US banks block third-party AI agents after detecting unauthorized account actions

    Major United States financial institutions, including JPMorgan Chase, Citigroup, and Bank of America, are blocking third-party artificial intelligence agents after detecting automated bots executing actions inside customer accounts without explicit user commands.

  12. OpenCode Console breach may have exposed users' OpenAI, Anthropic and Google keys

    OpenCode says an attacker broke into one of its systems and may have exposed the OpenAI, Anthropic and Google API keys that some users saved in OpenCode Console. The breach hit 1.4% of Console users, all of whom had an outside provider linked. OpenCode announced it on X on Octob…

  13. OpenAI mandates model router audits, Anthropic updates Claude API

    OpenAI announced that third-party AI gateways, model routers, and API resellers must pass formal compliance reviews to retain API access starting November 20, 2026. The policy introduces SafetyIDs, allowing intermediaries to isolate abusive accounts without exposing raw user ide…

    OpenAI

  14. Anthropic report shows junior developer hiring dropped to 9% as Claude Code expands

    Anthropic released an internal research report detailing the structural impact of AI agents on its engineering workforce. The share of junior engineers among new hires dropped from 27% to 9%, even as the company's overall headcount tripled over the same period.

    Anthropic

  15. Anthropic commits 5% of its shares to a new Public Benefit Fund

    Anthropic is putting 5% of its shares into the Anthropic Public Benefit Fund, a new wholly owned entity dedicated to the public good. The fund's stake can grow: for every 50% increase in Anthropic's value above a set baseline, the board can add up to another 2%. The first 5% sti…

    Anthropic

  16. OpenAI releases MultiLangBench: GPT-6 Astra passes 87.4% in rare languages

    OpenAI released MultiLangBench, a coding benchmark that asks models to build a SQLite-compatible database engine in 20 programming languages, including COBOL, Fortran, OCaml and Visual Basic. The test suite is on GitHub, and OpenAI also published results from a second evaluation…

    OpenAI

  17. Hottest AI startup funding rounds: Tenstorrent, Daytona, Arda, and Maxinsights

    Four big rounds this week: Tenstorrent (open RISC-V AI chips, led by Jim Keller), Daytona (sandboxed computers for AI agents), Arda Industries (AI that turns product ideas into CAD models and machine code) and Maxinsights (sensor-glove data to teach robots household chores).

  18. Top AI and tech startup rounds: Warhound, Warp Science, Slingshot, and Azoma

    Warhound raised $200 million at a $1.2 billion valuation for gene-edited military working dogs with Anduril gear. Warp Science raised $215 million for robot-run biology labs, Slingshot $27 million for AI micro-insurance at checkout, and Azoma $11.2 million to help brands get picked by AI shopping agents.

  19. The hottest AI startup funding rounds: Flapping Airplanes, Legora, Blacksmith, and Mach9

    Flapping Airplanes (biology-inspired reasoning models), Legora (AI agents for legal work), Blacksmith (fast test pipelines for code written by AI agents) and Mach9 (lidar scans to CAD maps in seconds) all raised big rounds this week.

  20. OpenAI acquires Orchestra Bio, a drug development agent platform

    OpenAI has acquired Orchestra Bio, a platform that helps biopharma teams coordinate drug development. Orchestra's team and technology will join Rosalind, OpenAI's AI models and tools for life sciences research.

    OpenAI

  21. Tencent releases Hy-Nomas, an open-source interactive world model runnable on an RTX 4090

    Tencent has released Hy-Nomas, an open-source foundation model for interactive video world generation that runs on consumer hardware down to a single NVIDIA GeForce RTX 4090. The release includes full training pipelines, inference code, interactive demos, and model weights under…

  22. Block announces Dessa, a full-duplex conversational model trained entirely on human speech

    Block has announced Dessa, an end-to-end full-duplex conversational speech model trained entirely on recordings of human conversation without text tokens or pretrained foundation models. Developed by a specialized research team reporting directly to Block Chairman Jack Dorsey, t…

  23. Inception releases Mercury 3, a diffusion LLM at 1,068 tokens per second

    Inception has released Mercury 3, a diffusion-based language model that runs at 1,068 tokens per second and has a 1M-token context window. It is available on the Inception API Platform and through OpenRouter, Baseten and Vercel.

  24. DeepSeek open-sources DeepInfer, its production inference engine

    DeepSeek has open-sourced DeepInfer, the inference framework that runs its own production services. The code is on GitHub under the Apache 2.0 license. With it, DeepSeek also released deepseek-recipe under the MIT license: the layer that turns API requests into model prompts and…

  25. DeepSeek opens Search and Fetch APIs at $0.75 per 1,000 requests

    DeepSeek has opened its Search and Fetch APIs to all developers and will charge $0.75 per 1,000 successful requests for each. Calls made through DeepSeek Harness, Claude Code and Codex are free for now. Billing for those starts on November 1, 2026, at 00:00 Beijing time (October…

  26. xAI releases Grok Voice 3.0, a voice agent model at $0.08 a minute

    xAI has released Grok Voice 3.0, a speech-to-speech model for real-time voice agents. It is available now in xAI's Voice Agent API as grok-voice-3.0 at $0.08 per minute, the same price as Grok Voice 2.0.

  27. Mistral launches Timestral, small forecasting models that run on a CPU

    Mistral has released Timestral, its first family of time-series foundation models. They make probabilistic, multivariate forecasts and run on a CPU with no GPU. Mistral announced the models on October 15, 2026.

  28. Exa launches Exa 3.0 search with a cheaper Lite tier

    Exa has released Exa 3.0, a new version of its web search engine for AI agents, with a larger embedding model, Dynamic Highlights now generally available and a bigger index. It is available today through the Exa API, along with a new cheaper search type called Exa Lite.

  29. Cloudflare previews Clairic, a typed data extraction model on Workers AI

    Cloudflare has released Clairic, a model that pulls typed values out of text and images instead of generating text. It is in preview on Workers AI as @cf/cloudflare/clairic, callable from a Worker or through the REST API.

  30. Dream Labs opens Dream-1, a model that drives robots it never trained on

    Dream Labs, a startup founded by four former NVIDIA GR00T researchers, has come out of stealth with Dream-1, a model that controls two-armed robots it was never trained on. Its public API is open now, from $3 per hour.

  31. RUMOR: Anthropic readies a new model, possibly Fable 5.5

    The Information reports that Anthropic is preparing a new model that people inside the company have already tested. The report says it should come out this month.

    Anthropic

  32. Coming in November: How LLMs work, a 23-lesson video course

    Our new video course, How LLMs work, comes out in November. It goes from the first artificial neuron to building your own LLM, in 23 animated lessons with hosts Anna and Max. Along the way, Akami builds his own chatbot, AkamiGPT, one piece in every lesson.