https://zubnet.ai/news/en/ AI platform news by Sarah Chen. en Fri, 31 Jul 2026 17:44:33 +0000 https://zubnet.ai/sarah.png https://zubnet.ai/news/en/ https://zubnet.ai/news/en/databricks-omnigent-open-source-meta-harness-compose-govern-agents/ https://zubnet.ai/news/en/databricks-omnigent-open-source-meta-harness-compose-govern-agents/ Databricks open-sourced Omnigent (Apache 2.0, alpha), a meta-harness, a layer that sits above individual agent harnesses like Claude Code, OpenAI Codex, and Pi and treats each as an interchangeable part. Its premise: engineers now juggle four or five agents at once, copy-pasting between them, and each harness only understands its own sessions. Omnigent adds a shared layer for composition (swap agents with one-line changes, /model mid-session), control (stateful YAML policies enforced at the meta-harness layer, not in a prompt, pause after $100 spent, require human approval before a git push that follows an npm install), and collaboration (share a live session by URL to watch, comment, co-drive, or fork). The part worth noting is where the guardrails live: outside the model, with its Omnibox sandbox keeping secrets out of the agent's reach, a GitHub token is never shown to the agent and is injected only by the egress proxy on approved requests. That is almost exactly the OWASP prescription from this week, bound the blast radius around a model you cannot trust to police itself, shipped as a tool. Sun, 14 Jun 2026 05:01:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/moonshot-kimi-k2-7-code-open-coding-model-21-percent-over-k2-6/ https://zubnet.ai/news/en/moonshot-kimi-k2-7-code-open-coding-model-21-percent-over-k2-6/ Moonshot AI released Kimi K2.7-Code, an open-weight (Modified MIT) coding model: a 1-trillion-parameter Mixture-of-Experts that activates about 32B per token, with MLA attention, a MoonViT vision encoder, native INT4, and a 256K context, weights (~595GB) on Hugging Face plus an OpenAI-compatible API. It is the model sibling of yesterday's Kimi Work desktop agent, which ran on the prior K2.6. The headline is +21.8% over K2.6 on Moonshot's own Kimi Code Bench v2 (50.9 to 62.0), and the fuller table is unusually candid: it improves on K2.6 across the board but mostly still trails GPT-5.5 and Claude Opus 4.8, with one flagged exception, it beats Opus 4.8 on the MCP Mark Verified tool-use benchmark (81.1 vs 76.4). It also uses about 30% fewer reasoning tokens than K2.6 (less overthinking) and undercuts the frontier on price ($0.19 cached / $0.95 miss input, $4 output vs Opus's $5/$25). Caveats: all numbers are first-party, competitors were tested in different environments, and 595GB needs server-class hardware. Sat, 13 Jun 2026 04:57:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/moonshot-kimi-work-local-desktop-agent-kimi-k2-6-300-sub-agent-swarm/ https://zubnet.ai/news/en/moonshot-kimi-work-local-desktop-agent-kimi-k2-6-300-sub-agent-swarm/ Moonshot AI shipped Kimi Work, a downloadable desktop app (macOS Apple silicon and Windows) that runs an AI agent locally: it reads your local files, drives a real browser with your logged-in sessions through a WebBridge extension, runs scheduled cron jobs, and fans a task out to as many as 300 parallel sub-agents, one per file in document triage, for instance, then merges. It reportedly runs on Kimi K2.6, the open-weight MoE that activates about 32B parameters per token with a 256K context, though the article flags that Moonshot did not officially confirm the model. The pitch is local-first: your machine, your folders, your authenticated browser, on a schedule, the same own-it posture as yesterday's Nous Hermes. The flip side, named in the product's own comparison, is that security responsibility moves to you, and a local agent with your files, your logged-in browser, and a 'YOLO mode' that drops the approval gate is the lethal trifecta pointed at your real accounts. No benchmarks, no pricing, and no MCP support are mentioned. Fri, 12 Jun 2026 09:37:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/moonshot-kimi-work-local-desktop-agent-300-sub-agent-swarm-k2-6/ https://zubnet.ai/news/en/moonshot-kimi-work-local-desktop-agent-300-sub-agent-swarm-k2-6/ Moonshot AI, the lab behind Kimi, shipped Kimi Work, a downloadable desktop app for macOS and Windows that runs AI agents locally on your machine instead of in the cloud. You give it a natural-language goal and it acts on your real files and browser. It runs on Kimi K2.6, an open-weight MoE with about 32B active parameters and a 256K context, and the headline feature is a swarm: up to 300 sub-agents in parallel on your own hardware, splitting a task and coordinating results, multi-agent orchestration on a laptop rather than a data center. Around it sit a WebBridge browser extension that drives your real browser using your existing logins, a cron scheduler with a keep-awake toggle for overnight runs and shell and Python execution, and direct local file access. It is the local-first posture again, your machine, your data, your keys. The catch is that the same capability is the lethal trifecta on your desktop, and the app is proprietary even though the K2.6 model is open. Fri, 12 Jun 2026 09:37:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/zyphra-zamba2-vl-hybrid-mamba2-transformer-vlm-10x-lower-time-to-first-token/ https://zubnet.ai/news/en/zyphra-zamba2-vl-hybrid-mamba2-transformer-vlm-10x-lower-time-to-first-token/ Zyphra released Zamba2-VL, an open vision-language model family in three sizes (1.2B, 2.7B, 7B) under Apache 2.0, and the architecture is the story. The language backbone alternates Mamba2 state-space layers, which carry the bulk of the compute and scale near-linearly with sequence length, with shared transformer attention blocks (each carrying a unique LoRA adapter) that preserve the in-context retrieval pure-SSM models give up. The payoff is latency on long visual input: a single high-res image can add roughly 3,400 vision tokens, and on a 32k-token prefill Zyphra reports about an order-of-magnitude lower time-to-first-token than the closest Transformer baseline, because the state-space backbone keeps a fixed-size recurrent state instead of a growing attention KV cache. The tradeoff is honest and printed in the benchmarks: strong on visual counting and document understanding, behind same-size Transformers on knowledge reasoning and OCR. It needs a CUDA GPU, self-hosting, and ships as a research artifact, not an API. Fri, 12 Jun 2026 08:06:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/jeff-bezos-prometheus-12b-artificial-general-engineer-physical-world/ https://zubnet.ai/news/en/jeff-bezos-prometheus-12b-artificial-general-engineer-physical-world/ Prometheus, the physical-AI startup co-founded by Jeff Bezos and ex-Verily co-founder Vik Bajaj, raised a $12 billion Series B at a $41 billion valuation, after a $6.2 billion Series A late last year. Its pitch is an artificial general engineer: software meant to automate the design and manufacturing of complex physical systems, from jet engines to drug compounds. Note the framing, not AGI but a general engineer, which puts it at a different layer of the physical-AI stack than the robots and world-simulators we keep covering, the engineering brain that designs the artifact and the process to build it. The investor list is the tell: alongside Bezos, the backers are JPMorgan, Goldman Sachs, and BlackRock, Wall Street balance sheets rather than only venture capital, the way you fund something compute-heavy. The honest gap is between the price and the proof: $41 billion and 150 employees, with the artificial general engineer described as a goal, not a demonstrated capability, and no shipped product, benchmark, or engine in the reporting. Fri, 12 Jun 2026 01:04:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/visa-chatgpt-ai-agents-retail-purchases-single-use-payment-tokens/ https://zubnet.ai/news/en/visa-chatgpt-ai-agents-retail-purchases-single-use-payment-tokens/ Visa connected its payment network to ChatGPT so an AI agent can not just recommend a product but complete the purchase, no human at the final click, at supporting merchants. Per the report, the mechanism is programmatic tokenization: you pre-authorize the ChatGPT environment with spending parameters, and when the agent decides to buy it generates a single-use payment token through the Visa network that hits the merchant backend like a digital-wallet payment with the visual checkout removed. The agent shops on machine-readable inventory, not storefronts, the same agent-readable-web shape as WebMCP, applied to money. The catch is the timing: it lands the same day as an OWASP report naming prompt injection the top agentic failure, and the article itself concedes injection could steer an agent to a malicious vendor. An agent with a payment token, untrusted merchant data, and the power to transact is the lethal trifecta lit up, except what leaks is your money. Single-use tokens and spending caps bound the blast radius; they do not close it. The report names no Visa product, no limits, and no clear live-vs-pilot status. Thu, 11 Jun 2026 11:08:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/nous-research-hermes-agent-profile-builder-identity-model-skills-mcp/ https://zubnet.ai/news/en/nous-research-hermes-agent-profile-builder-identity-model-skills-mcp/ Nous Research gave its open-source Hermes Agent a web dashboard that turns a string of CLI commands into one guided flow for building an isolated agent profile. Launch hermes dashboard, it opens on localhost only, and you fill in five groups: identity (the name becomes a shell alias, so a profile called coder answers to coder chat), model and provider (Nous Portal, OpenRouter, NVIDIA, OpenAI, or any OpenAI-compatible endpoint), built-in skills, Skills Hub installs, and MCP servers (HTTP by URL or stdio by command). Each profile is its own home directory: a config.yaml, an .env for keys, a SOUL.md for personality, plus separate memory, sessions, skills, cron, and state. The point is isolation, a coding agent and a research agent never share state, and the unit it formalizes is the profile: an agent as a portable, named bundle of identity, model, tools, memory, and a soul file you run locally. Thu, 11 Jun 2026 09:53:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/google-diffusiongemma-text-diffusion-4x-faster-lower-quality-open-model/ https://zubnet.ai/news/en/google-diffusiongemma-text-diffusion-4x-faster-lower-quality-open-model/ Google DeepMind released DiffusionGemma, an experimental open model that generates text the way diffusion models generate images: instead of one token at a time, left to right, it denoises a whole block of up to 256 tokens per step, refining from random placeholders with bi-directional attention. It is a 26B Mixture of Experts (3.8B active) built on Gemma 4, Apache 2.0, weights on Hugging Face, fits in 18GB quantized, and runs about 4x faster in single-user use: 1000+ tokens per second on an H100, 700+ on an RTX 5090. The catch is stated plainly by Google: overall output quality is lower than standard Gemma 4, which it recommends for maximum quality. The deeper point is the deployment inversion: diffusion shifts the decode bottleneck from memory bandwidth to compute, so it wins on local, low-concurrency hardware, the opposite of cloud serving economics. Wed, 10 Jun 2026 16:24:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/decart-oasis-3-world-model-photorealistic-driving-collision-physics/ https://zubnet.ai/news/en/decart-oasis-3-world-model-photorealistic-driving-collision-physics/ Decart launched Oasis 3, a real-time interactive world model that generates photorealistic driving environments from a text prompt and serves three camera views (one front, two side) for autonomous-vehicle training, via API at $0.02 per second. CEO Dean Leitersdorf says it is more than an order of magnitude cheaper to run than anything else in the field. The honest part is in the headline: over long runs the world degrades, locations do not persist, control slips, and crucially the model does not simulate collisions, cars drive straight through each other. That last gap is exactly what AV training needs most, and Decart calls it a major research problem it is cracking now. World models just crossed from research demo into priced commercial infrastructure, with the caveats printed on the box. Wed, 10 Jun 2026 13:07:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/azure-api-management-mcp-content-safety-unified-model-api-ai-gateway/ https://zubnet.ai/news/en/azure-api-management-mcp-content-safety-unified-model-api-ai-gateway/ At Build 2026, Azure API Management shipped two AI-gateway pieces. A Unified Model API (public preview) lets clients write once in OpenAI Chat Completions format while APIM translates to Anthropic, Google Vertex, and other backends, so you swap or route across providers without touching client code, with governance applied centrally. The sharper piece extends APIM's content-safety policy beyond LLM text to Model Context Protocol tool calls and Agent-to-Agent payloads, scanning tool-call arguments, MCP responses, and inter-agent messages, with category filters and prompt-injection detection. It lands the same week the gateway is the thing under attack: the LiteLLM command-injection bug on CISA's active-exploitation list is the other face of the same realization, that if all model and agent traffic flows through one control plane, that plane is both where you get attacked and where you enforce policy. Wed, 10 Jun 2026 09:38:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/wwdc-2026-apple-siri-gemini-teacher-afm-distilled-private-cloud-compute-ios-27/ https://zubnet.ai/news/en/wwdc-2026-apple-siri-gemini-teacher-afm-distilled-private-cloud-compute-ios-27/ Apple's WWDC 2026 keynote confirmed the shape of the reported roughly $1B-a-year Google deal: the new Siri is built with Gemini, but as teacher, not tenant. Apple Foundation Models are distilled from loaned Gemini models and rebuilt for Apple silicon, AppleInsider's sources insist 'not a drop of Gemini' ships in the weights, and heavy requests route to AFM Cloud on Private Cloud Compute, where Apple says data is used only to execute the request, never stored, and externally verifiable. Around it: systemwide AI dictation, natural-language Shortcuts, cross-app context, Photos Reframe and Extend, iOS 27 on iPhone 11 and newer, and foldState APIs hinting at September hardware. Tue, 09 Jun 2026 18:04:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/anthropic-claude-fable-5-mythos-class-goes-public-fallback-opus-4-8-safety/ https://zubnet.ai/news/en/anthropic-claude-fable-5-mythos-class-goes-public-fallback-opus-4-8-safety/ Anthropic released Claude Fable 5 on June 9, the first generally available model from the Mythos class, the capability tier above Opus that stayed gated behind Project Glasswing since April. The safety design is new: instead of refusing high-risk requests in offensive cyber, biology, or distillation, classifiers fall the session back to Opus 4.8, in under 5% of sessions on average. Pricing is $10 per million input tokens, $50 per million output. The receipts include Stripe compressing a 50-million-line Ruby migration from two months to a day. A restricted sibling, Claude Mythos 5, ships the same weights with safeguards selectively lifted for vetted partners. Capability access just became a graduated privilege. Tue, 09 Jun 2026 17:00:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/webmcp-chrome-149-origin-trial-websites-declare-tools-browser-agents/ https://zubnet.ai/news/en/webmcp-chrome-149-origin-trial-websites-declare-tools-browser-agents/ WebMCP moved from a behind-a-flag prototype to a public Origin Trial in Chrome (149 through 156), the Chrome team announced. It is a proposed open web standard that lets a website declare structured tools, JavaScript functions through an Imperative API or annotated HTML forms through a Declarative API, that an in-browser AI agent can call directly with document.modelContext.getTools() and executeTool(). The pitch over today's screenshot-and-click vision agents: one typed call instead of a pixel-hunting loop, with early benchmarks citing 8 to 12 times faster task completion, and the agent can only invoke tools the site explicitly exposed. The catch worth naming: the only consumer today is Gemini in Chrome, so a web standard currently has one client. Tue, 09 Jun 2026 15:39:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/google-gemini-3-5-live-translate-real-time-speech-to-speech-70-languages/ https://zubnet.ai/news/en/google-gemini-3-5-live-translate-real-time-speech-to-speech-70-languages/ Google DeepMind shipped Gemini 3.5 Live Translate, a speech-to-speech model that translates across 70+ auto-detected languages a few seconds behind the speaker while preserving the original voice's intonation, pacing, and pitch. It lands in three places at once: the Live API in public preview (gemini-3.5-live-translate-preview), Google Meet enterprise preview this month, expanding translated speech from 5 languages to 70+ and 2,000+ pairs, and the Google Translate app on Android and iOS. Launch partners include Agora, LiveKit, Pipecat, and Grab, which is testing it on 10M+ monthly voice calls. Every second of output audio carries a SynthID watermark. Not published: pricing, and any quantitative quality metric. Tue, 09 Jun 2026 15:16:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/lovable-500m-annualized-revenue-1-million-projects-week-vibe-coding/ https://zubnet.ai/news/en/lovable-500m-annualized-revenue-1-million-projects-week-vibe-coding/ Lovable reports $500M in annualized run-rate revenue as of June 2026, up from $400M in February, with 1 million new projects created every week and 50 million built since launch. The vibe-coding platform, founded in late 2023, lets non-technical users, founders, designers, salespeople, build storefronts, CRMs, and internal tools by describing them. TechCrunch frames it inside the so-called SaaSpocalypse: why pay for an expensive SaaS contract when you can describe your own tool into existence. No churn or profitability figures were disclosed, and weekly project creation is not retained software. But the demand-side signal is unambiguous. Tue, 09 Jun 2026 13:00:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/microsoft-discovery-ga-agentic-science-platform-majorana-2-quantum-chip/ https://zubnet.ai/news/en/microsoft-discovery-ga-agentic-science-platform-majorana-2-quantum-chip/ Microsoft Discovery, a platform for running teams of specialized AI agents through scientific R&D loops, reached general availability on Azure. The showcase is concrete in a way benchmarks are not: Microsoft credits Discovery's agents with central work on the Majorana 2 quantum chip, managing fabrication, automating measurements, optimizing the materials stack, and correlating patterns across two decades of experimental data to surface manufacturing flaws that had gone unnoticed. Microsoft says the result was a 1,000-fold reliability gain, a mean qubit lifetime of 20 seconds, and a timeline pulled from 2034 to 2029. These are the vendor's numbers about the vendor's chip, so the honest frame is that the agents were a credited instrument, not an independently audited cause. The design detail that travels: confidence scoring and citations, so outputs come with a checkable evidence trail. Mon, 08 Jun 2026 09:08:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/xai-grok-build-0-1-api-mcp-native-agentic-coding-model/ https://zubnet.ai/news/en/xai-grok-build-0-1-api-mcp-native-agentic-coding-model/ xAI opened grok-build-0.1 on its public API, the same model that powers the Grok Build CLI, now usable without a SuperGrok or X Premium+ subscription. It is trained from the ground up for agentic coding, planning a task, executing it step by step, debugging failures, and calling tools through MCP, with a 256K context window, output served at 100-plus tokens per second, and pricing of $1 per million input tokens and $2 per million output. The signal here is access, not a new capability ceiling: a model that was gated behind a consumer subscription is now a raw, tool-fluent API priced for builders. No independent agentic-coding benchmark ships with the announcement, so the codes-like-an-agent claim is vendor-stated until your own harness says otherwise. Fri, 05 Jun 2026 01:37:00 +0000 sarah@zubnet.ai (Sarah Chen) https://zubnet.ai/news/en/google-antigravity-2-0-adds-cli-sdk-managed-agents-tier/ https://zubnet.ai/news/en/google-antigravity-2-0-adds-cli-sdk-managed-agents-tier/ Google shipped Antigravity 2.0 at I/O 2026 as a 5-surface platform, the desktop IDE from version 1.0 plus a new CLI, an SDK for custom agents, a Managed Agents runtime tier inside the Gemini API, and an enterprise path via Gemini Enterprise Agent Platform. Powered by Gemini 3.5 Flash, co-developed using Antigravity itself. The pitch to builders: stay in one substrate from local development to production agent deployment. Tue, 19 May 2026 17:00:00 +0000 sarah@zubnet.ai (Sarah Chen)