Today’s lead · x.com
Gemini 3.5 Flash adds native computer use for browser, mobile, and desktop agents
Google DeepMind says computer use is now a built-in tool in Gemini 3.5 Flash, moving the capability from a standalone Gemini 2.5 computer-use model into the main Flash API and Gemini Enterprise Agent Platform. The model can see, reason, and act across browser, mobile, and desktop interfaces for tasks such as software testing and knowledge-work automation. Google is also adding optional enterprise safeguards for explicit confirmation on sensitive actions and automatic stopping when indirect prompt injection is detected, while still recommending sandboxing, human review, and strict access controls.
Top signals
14 more
Report: White House asks OpenAI to limit GPT-5.6 access before broader release
TechCrunch, citing The Information, reports that OpenAI plans to release GPT-5.6 first to a small set of partners after the White House asked for a slower rollout over safety concerns. CEO Sam Altman reportedly told staff that the government would approve access customer by customer during a preview period, with OpenAI hoping for a broader release a couple of weeks later if the limited launch goes well.
Gaslight macOS malware targets AI-assisted reverse engineering with prompt injection
SentinelOne researchers disclosed Gaslight, a previously undocumented Rust-based macOS implant and infostealer attributed with high confidence to North Korea-aligned actors. The malware uses Telegram for command-and-control, installs persistence through a LaunchAgent, drops a Python stealer for Keychain, browser, process, hardware, and shell-history data, and embeds 38 fabricated “system” messages intended to make LLM-assisted triage agents abort or distrust their analysis.
OpenRouter pushes live model selection into agents and team chat stacks
OpenRouter introduced an MCP server that gives agents live model intelligence for choosing, pricing, and testing models instead of relying on stale training-time knowledge. It also added OpenWebUI support for one chat interface and unified billing across 400+ models, while GLM-5.2 availability expanded through OpenRouter’s “z-ai/glm-5.2:nitro” fast-provider route and Cursor support via Fireworks.
Sail Research raises for long-running agent inference
Sail Research is building an inference platform for long-running agents and has raised an $80 million Seed and Series A round, including a $20 million investment from Redpoint alongside Kleiner Perkins, Sequoia, Theory, and others. The technical pitch is an agent-specific inference stack across heterogeneous hardware, custom kernels, and primitives for pausing, resuming, and forking whole VMs, including open network connections.
Agent evaluation bets shift toward simulated worlds and gameplay-derived action data
Patronus AI raised a $50 million Series B, bringing total funding to $70 million, to build “digital worlds” that replicate websites and internal systems for stress-testing agents after reinforcement learning in domains like software engineering and finance. General Intuition separately raised $320 million at a $2.3 billion valuation to train agentic models on gameplay clips with action labels from Medal, arguing that button-press data can transfer from games and simulation toward robotics and real-world spatial-temporal reasoning.
Unconventional AI demos Un-0 on a simulated oscillator computer
Unconventional AI, led by former Databricks AI chief Naveen Rao, released Un-0, an image-generation system built on a software simulation of its oscillator-based computing architecture. The company says the approach can match state-of-the-art diffusion-style outputs today and could eventually cut AI inference power use by up to 1,000x once real chips and a full serving stack exist.
Agent tooling spreads into persistent dev boxes, WASM sandboxes, and PowerPoint CLIs
OpenAI Devs says Codex can now create a persistent DigitalOcean cloud development environment from one prompt, running inside the user’s DigitalOcean account and continuing while they step away. Rivet released agentOS v0.2.0, a Rust-rewritten WebAssembly sandbox and orchestration layer for multiple agent harnesses, while Hands-on Deck exposes PowerPoint-style inspect, edit, render, and diff operations as a CLI so agents can make atomic .pptx changes instead of hand-editing OOXML.
Amazon commits another $13B to AI and cloud infrastructure in India
Amazon says it will invest an additional $13 billion through 2030 to expand AWS data center capacity in Mumbai and Hyderabad, bringing its stated India commitments to $48 billion. The announcement follows major India infrastructure commitments from Microsoft and Google and reinforces India’s role as a target region for AI compute expansion.
Netris raises $15M to automate neocloud networking for GPU clusters
Netris raised a $15 million Series A from Andreessen Horowitz for switch-level automation that helps AI neoclouds configure, operate, and isolate GPU cluster networks faster. The company says its vendor-agnostic platform is already live across more than 35 GPU clusters totaling about one million GPUs, with support for Nvidia and AMD server environments.
New audio and video APIs target faster multimodal production
fal says Seed Audio 1.0 is now exclusively available on its platform, generating voice, music, and sound effects in one pass while accepting up to three reference audio clips and building multi-speaker scenes from a prompt. BytePlus says ByteDance’s Dreamina Seedance 2.0 Mini API is live for developers who need faster video generation at scale.
Adobe buys Topaz Labs to deepen AI image and video enhancement
Adobe is acquiring Topaz Labs, maker of AI image and video enhancement tools including Astra for video upscaling and Wonder for image retouching. Adobe plans to integrate Topaz models into Firefly and its creative apps while keeping Topaz products available as standalone services, with the transaction expected to close in the second half of 2026.
Step 3.7 Flash gets a coding-agent demo and startup push
A Step 3.7 Flash demo shows the model generating a playable self-contained Flappy Bird-style HTML game from one prompt in under 30 seconds, including canvas gameplay, gravity, collision detection, scoring, game-over state, and restart behavior. The linked product page positions the model for production agents with multimodal input, web and visual search, tool calling, and compatibility with harnesses such as Claude Code, KiloCode, Hermes Agent, OpenClaw, MCP, and Skills.
NVIDIA says Cosmos 3 Nano tops Allen AI’s MolmoSpaces robot policies leaderboard
NVIDIA Robotics says a policy model post-trained on Cosmos 3 Nano ranks No. 1 on Allen AI’s MolmoSpaces robot policies leaderboard. The benchmark focuses on environments where models must interpret, reason, and act in real time, making the result more relevant to embodied-agent policy work than static vision-language scoring alone.
DATA Network launches Trace for dataset provenance audits
DATA Network introduced Trace, a public audit and search platform for assets permanently registered on its network. AI labs and regulators can search human-contributed photos or voice samples and filter by dataset, app, data type, modality, and time to inspect an individual record’s provenance trail.
Tools & repos
3 selectedgoogle-labs-code/design.md
A format specification for describing a visual identity to coding agents. DESIGN.md gives agents a persistent, structured understanding of a design system.
calesthio/OpenMontage
World's first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio.
xbtlin/ai-berkshire
AI 时代的伯克希尔：基于 Claude Code 的价值投资研究框架。巴菲特·芒格·段永平·李录四大师方法论 + 多Agent并行研究。| AI-era Berkshire: a value investing research framework built on Claude Code. 4 masters' methodologies + multi-agent adversarial analysis.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Funding & acquisitions
5 movesOxlo.ai
Scale across AI models without scaling your bill
BrowserAct
Web browser automation for AI agents
Zaro
Build agents & apps on top of your context with one prompt.
Figma Motion
Your Figma canvas now has a timeline
Brain² by ClickUp
One AI that knows your entire company and acts on it
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.