Latest AI News

Anthropic is expanding Project Glasswing, its AI-powered cybersecurity initiative, from roughly 50 initial partners to approximately 150 organizations across more than 15 countries. Partners use Claude Mythos Preview to scan codebases for vulnerabilities, having already found over 10,000 high- or critical-severity flaws. The new cohort covers critical infrastructure sectors including power, water, healthcare, and communications. Anthropic estimates a successful attack on most partners' codebases could affect more than 100 million people, underscoring the program's national and global security stakes.

Perplexity has announced a hybrid local-server inference orchestrator for its Personal Computer product, launching in July 2026. The system automatically routes tasks between on-device models and cloud-based frontier models, keeping sensitive data like financial records and health information local while sending compute-intensive work to servers. Unveiled in partnership with Intel, it also supports NVIDIA RTX Spark silicon. Perplexity argues the architecture reduces centralized infrastructure demand and improves data sovereignty without requiring countries to build dedicated data centers.

Anthropic has added dynamic multi-agent workflows to Claude Code, allowing the AI to write and orchestrate its own custom harness on the fly for complex tasks. Unlike static workflows, dynamic workflows spawn separate Claude subagents with isolated context windows, combating failure modes like agentic laziness, self-preferential bias, and goal drift. Powered by Claude Opus 4.8, workflows execute JavaScript files to coordinate subagents and support patterns like fan-out-and-synthesize, adversarial verification, and tournament-style competition. Users can trigger them by asking Claude or using the keyword "ultracode."

President Trump signed an executive order on June 2, 2026, directing federal agencies to strengthen AI cybersecurity within 30 days. CISA must issue binding directives to harden civilian systems and expand AI-enabled defensive tools. The Treasury Department will form a voluntary AI cybersecurity clearinghouse with industry to scan and patch software vulnerabilities. A separate voluntary framework will allow AI developers to submit frontier models for classified government review up to 30 days before release. The order explicitly prohibits mandatory licensing or preclearance requirements for AI models.

OpenAI is expanding Codex, now used by over 5 million people weekly, with role-specific plugins, shareable interactive sites, and inline annotations. Six new plugins cover data analytics, creative production, sales, product design, public equity investing, and investment banking, bundling 62 apps and 110 skills. Non-developers make up 20% of Codex users and are growing three times faster than developers. Sites, rolling out in preview for Business and Enterprise customers, let teams share interactive dashboards and apps via URL.

Microsoft has unveiled Majorana 2, a scalable quantum processor featuring topological qubits that are over 1,000 times more reliable than those in its predecessor. By replacing aluminum with lead in the material stack, qubit lifetimes jumped from 1–12 milliseconds to a mean of 20 seconds, occasionally exceeding one minute. The larger topological gap better shields qubits from environmental noise. AI assisted in designing the new stack. Microsoft, one of two finalists in DARPA's US2QC program, has halved its roadmap timeline and now targets a fault-tolerant quantum computer by 2029.

Microsoft has announced the private preview of Frontier Tuning, a new AI customization approach designed to make AI models work according to specific business needs. The system applies reinforcement learning inside a company's compliance boundary, using the organization's own data, processes, and conventions. This allows businesses to tailor AI behavior without exposing sensitive information outside their controlled environment, making it particularly relevant for enterprises with strict data governance requirements.

Microsoft has introduced Scout, its first Autopilot agent for Microsoft 365, a new category of always-on agents that work autonomously on your behalf without needing to be prompted each time. Scout integrates with Teams, Outlook, OneDrive, and SharePoint, scheduling meetings, blocking calendar time for deliverables, and flagging risks like stalled decisions. It uses Work IQ to learn user priorities over time and is powered by OpenClaw open-source technology. Scout is currently in private preview for Frontier organizations requiring a GitHub Copilot license.

Microsoft held its Build 2026 developer conference, using the annual event to connect with the global developer community and highlight progress on Windows as a trusted development platform. The company described Build as one of its favorite moments each year for sharing what its teams have been building. However, the available source text is truncated, limiting full detail on specific announcements, features, or tools revealed at the event.

Microsoft will make its Work IQ APIs generally available on June 16, 2026, offering developers a new way to build AI agents that interact with Microsoft 365 data. Work IQ acts as a semantic intelligence layer, continuously processing email, calendar, meetings, chats, files, and organizational data to give agents business context rather than raw data. The APIs claim 2x faster processing, 80% fewer tokens than traditional APIs, and collapse tool calling into just 10 generic tools via MCP. Pricing uses consumption-based Copilot Credits.

Microsoft has launched Web IQ, a suite of AI-native grounding APIs built on Bing's global index but re-architected from the ground up for agentic AI workloads. Web IQ connects AI agents to fresh web data including pages, news, images, and videos, returning passage-level evidence rather than full documents to improve token efficiency. It operates at sub-165ms p95 latency, nearly 2.5 times faster than competitors, and outperforms alternatives on grounding satisfaction scores measured across 3,000 global queries.

Microsoft has unveiled Project Solara, a chip-to-cloud software platform paired with tailored hardware designed to enable agent-first devices. Built on the premise that the next platform shift moves from apps to agents, Solara treats agents as both a new programming unit and a new human-to-machine interaction model. The platform supports an open, multi-agent world with enterprise-grade security, identity, and manageability built in. Microsoft is previewing two device concept categories—stationary and portable—targeting verticals including healthcare, retail, and finance.

Microsoft has announced the Surface RTX Spark Dev Box, a new device aimed at software developers. Revealed on the Windows Devices blog, the machine is designed to meet the demanding workloads of modern development, including longer-running tasks and intensive tooling requirements. Microsoft positions the device as built to match the pace of contemporary software creation, though specific hardware specifications, pricing, and availability details were not fully captured in the available source text.

GitHub has launched the Copilot app, a desktop control center for agent-native development, now in technical preview for existing Copilot Pro, Pro+, Business, and Enterprise users. The app features a My Work view to manage parallel agent sessions, pull requests, and background automations across connected repositories. New canvases provide bidirectional work surfaces where agents and developers collaborate visibly. Additional features include cloud and local sandboxes, an Agent Merge tool, upgraded code review with medium-tier reasoning models, and a generally available Copilot SDK supporting Node.js, Python, Go, .NET, Rust, and Java.

Mayo Clinic and Microsoft have announced a strategic collaboration to develop a frontier AI model purpose-built for healthcare. The model will combine Mayo Clinic's de-identified clinical health data and longitudinal insights with Microsoft's AI and cloud capabilities to support clinical reasoning, earlier diagnoses, and personalized treatment decisions. Mayo Clinic will own the model, reinforcing patient trust and data stewardship. Microsoft plans to distribute it via Azure Foundry APIs. Microsoft AI CEO Mustafa Suleyman called it the best collaboration imaginable to accelerate frontier medical intelligence.

Black Forest Labs has named legendary filmmaker Martin Scorsese as an advisor, bringing his six decades of cinematic storytelling experience to the AI image generation company. Scorsese will help guide the development of visual intelligence with human taste and craft as central priorities. The announcement was accompanied by a working storyboarding session in which Scorsese used Black Forest Labs' FLUX image generation model, demonstrating a practical creative collaboration between the director and the AI tool.

OpenAI broke ground on The Barn, a 1GW Stargate data center campus in Saline, Michigan, alongside Governor Gretchen Whitmer and partners Oracle, Related Digital, and Walbridge. The project is expected to create 2,500 union construction jobs and 450 permanent onsite positions, generating $1 billion in tax revenue over the lease term. OpenAI will also provide up to $45 million in Codex credits to over 400,000 eligible Michigan college and trade school students during the 2026–2027 academic year.

TwelveLabs has launched Rodeo, its first application-layer product and an AI-powered creative copilot that lets video editors find, assemble, and edit footage using natural language. Powered by TwelveLabs' Marengo 3.0 and Pegasus 1.5 foundation models, Rodeo searches entire footage libraries contextually, understanding not just what appears on screen but why it matters. CEO Jae Lee says the tool marks TwelveLabs' evolution from infrastructure provider to creator-facing toolmaker, enabling producers and editors to go from raw footage to finished stories in minutes.

Microsoft's Build conference in San Francisco will feature several major announcements, including a new reasoning model called MAI-Thinking-1 from Microsoft AI chief Mustafa Suleiman, plus MAI-Image-2.5 and MAI-Image-2.Flash. Microsoft will also unveil a developer-optimized Windows 11 experience with pre-installed tools and a distraction-free environment, a Copilot super app combining AI assistants into one interface, and expanded support for local AI models on Nvidia RTX Spark hardware. GitHub improvements and Qualcomm's continued Windows on Arm work are also expected.

JetBrains has released Mellum2, a 12B-parameter Mixture-of-Experts model trained from scratch on text and code. The model activates only 2.5B parameters per token, enabling more than 2x faster inference compared to similarly sized open models. Released under Apache 2.0, Mellum2 targets latency-sensitive workloads including routing, RAG pipelines, sub-agents, summarization, and private deployments. It is available on Hugging Face and is designed to serve as a fast, efficient component within larger multi-model AI systems for software engineering.

Draft summary: Anthropic, the AI safety company behind Claude, has confidentially submitted a draft S-1 registration statement to the U.S. Securities and Exchange Commission for a proposed IPO of its common stock. The filing gives Anthropic the option to go public once the SEC completes its review, though the offering remains subject to market conditions. Share count and pricing have not yet been determined. Anthropic recently raised $65 billion in Series H funding at a $965 billion post-money valuation.

OpenAI is actively hiring engineers for its robotics division, according to a post from CEO Sam Altman. The push signals OpenAI's growing ambitions in physical AI, aiming to develop robots capable of performing genuinely useful real-world tasks. While specific roles and technical details were not disclosed in the post, the recruitment drive suggests OpenAI is scaling its robotics team significantly as competition in humanoid and general-purpose robotics intensifies across the AI industry.

Microsoft has unveiled the Surface Laptop Ultra, its most powerful laptop ever, featuring an NVIDIA Blackwell RTX GPU with full CUDA support, up to 128GB of unified memory, and 1 petaflop of AI compute capable of running models up to 120 billion parameters locally. The inch mini-LED PixelSense Ultra touchscreen reaches 2,000 nits peak HDR brightness. Targeted at developers, creators, and AI builders, the device launches later in 2026.

NVIDIA has announced a major expansion of its DRIVE Hyperion level ready robotaxi platform at GTC Taipei, adding global partners across Asia, Europe, and the Middle East. Foxconn will deploy robotaxis in Kaohsiung, Taiwan starting in 2028, with airport-to-city routes planned. VinFast and Autobrains will bring level 4 vehicles to Southeast Asia. Uber and Autobrains will launch a robotaxi program in Munich. HUMAIN will deploy DRIVE Hyperion-powered robotaxis in Saudi Arabia.

NVIDIA announced DGX Station for Windows at GTC Taipei, a deskside AI supercomputer powered by the GB300 Grace Blackwell Ultra Desktop Superchip capable of running frontier AI models up to 1 trillion parameters locally. It features up to 748GB of coherent memory, 20 petaflops of FP4 performance, and 800Gb/s networking via ConnectX-8 SuperNIC. Developed with Microsoft, it supports NVIDIA OpenShell for secure agent sandboxing. Systems will be available from ASUS, Dell, HP, and others in Q4 2025.

NVIDIA announced Cosmos 3 at GTC Taipei at COMPUTEX, an open world foundation model combining vision reasoning, multimodal generation, and native action prediction to help robots, autonomous vehicles, and vision AI agents reason before acting. Using a mixture-of-transformers architecture, it generates outputs including synthetic video, joint angles, and trajectory points. Agile Robots uses it for humanoid task data, while Linker Vision applies it to smart city camera analysis. Cosmos 3 tops benchmarks including VANTAGE-Bench and Physics-IQ, and is available on Hugging Face and build.nvidia.com under the OpenMDW 1.1 license.

NVIDIA announced the Isaac GR00T Reference Humanoid Robot at GTC Taipei, an open humanoid robot reference design combining a Unitree H2 Plus chassis, Sharpa Wave tactile five-finger hands, and NVIDIA Jetson AGX Thor T5000 compute featuring a Blackwell GPU with 2,070 FP4 teraflops. The full-stack platform covers data capture, simulation, training, and deployment. Research institutions including Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego will use the design. The robot will be available from Unitree in late 2026.

NVIDIA has launched Vera, its first CPU designed specifically for AI agents and data center workloads, now in full production. Powered by 88 custom Olympus cores and LPDDR5X memory delivering up to 1.2TB/s bandwidth, Vera offers 1.8x faster task completion than x86 CPUs. It powers standalone Vera servers, Vera Rubin systems, and Vera BlueField-4 STX storage platforms. Adopters include Anthropic, OpenAI, Oracle Cloud Infrastructure, ByteDance, and CoreWeave, with systems from Dell, HPE, Lenovo, and Supermicro available this fall.

OpenAI's Codex app version 26.527 brings two significant Windows upgrades. Computer Use now works on Windows, enabling Codex to operate desktop applications by seeing, clicking, and typing directly in the foreground. Remote control support has also been extended to Windows devices, allowing users to access and control their Windows machines remotely through the ChatGPT mobile app on either iOS or Android. Together, these additions make Codex a more capable tool for Windows-based automation and remote workflows.

Perplexity has expanded its Computer AI agent into Microsoft Word, Excel, PowerPoint, and Outlook via native add-ins available on Microsoft Marketplace. The integration removes the copy-paste workflow that nearly 40% of Computer users already relied on when generating Microsoft-format outputs. Users can draft documents, build financial models, design presentations, and compose emails from a side panel, with data pulled from the web, SharePoint, FactSet, and Snowflake. Every figure is cited. The feature is available to Pro, Max, and Enterprise subscribers.

Anthropic has raised $65 billion in a Series H funding round led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money. The round includes $15 billion in previously committed hyperscaler investments, including $5 billion from Amazon. Anthropic's run-rate revenue crossed $47 billion earlier this month. Funds will advance safety and interpretability research, expand compute, and scale products. Strategic infrastructure partners Micron, Samsung, and SK hynix also joined the round.

Microsoft has redesigned the Microsoft 365 Copilot app and its in-app experience across Word, Excel, PowerPoint, and Outlook. The updated app loads more than 50% faster and improves complex chat response times by 10%. A new intelligence layer called Work IQ draws on emails, files, chats, and meetings to adapt responses. Since rolling out the redesign, Copilot usage rose 27% in Word, 33% in Excel, 43% in PowerPoint, and 30% in Outlook.

OpenAI has launched Rosalind Biodefense, a program giving trusted developers sponsored access to GPT-Rosalind, its frontier reasoning model for life sciences, to build biodefense and pandemic preparedness tools. OpenAI is also expanding GPT-Rosalind access to select U.S. government and allied partners for public health missions. Early partners include Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, the Coalition for Epidemic Preparedness Innovations supporting its 100 Days Mission, and Fourth Eon Biosecurity, which focuses on DNA synthesis screening.

OpenAI has published a playbook for conducting trustworthy third-party evaluations of frontier AI models, emphasizing that modern models require more than simple prompt-response testing. The guide highlights the critical role of the "harness" — the surrounding setup including tools, scaffolding, and context — which can significantly alter measured performance. OpenAI recommends evaluators specify claims being tested, document harness choices, and account for issues like reward hacking, sandbagging, and contamination. For example, GPT-5.5 cyber range scores improved materially when harness compaction was enabled.

Anthropic plans to widen release of its Claude Mythos cybersecurity AI model to all customers in the coming weeks, Reuters reported, after initially restricting it to Project Glasswing, a defensive program including Amazon, Microsoft, Apple, and Cloudflare. Mythos can identify software vulnerabilities, build exploit chains, and reason through attack paths previously requiring elite researchers. Cloudflare noted the model's safety refusals were inconsistent. Market pressure from rivals like OpenAI's GPT-5.5 and Mistral is accelerating Anthropic's timeline, with tiered access and pricing controls expected.

Anthropic has launched dynamic workflows in Claude Code, now in research preview, enabling Claude to write orchestration scripts that spin up tens to hundreds of parallel subagents in a single session. Designed for large-scale engineering tasks like codebase-wide bug hunts, framework migrations, and security audits, the feature is available via the CLI, Desktop, VS Code extension, and API, including Amazon Bedrock, Vertex AI, and Microsoft Foundry, for Max, Team, and Enterprise plans. Developer Jarred Sumner used it to port Bun from Zig to Rust—roughly 750,000 lines—in eleven days.

ElevenLabs has launched Dubbing v2, a new dubbing model designed to preserve the emotion and performance of original content across multiple languages. The company describes it as a revolutionary advancement, marking the first time dubbed audio can carry over the emotional nuance of a source performance into every target language. This release positions ElevenLabs as a stronger competitor in AI-powered localization and content translation for global media distribution.