AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Microsoft Unveils Seven In-House MAI Models at Build 2026, Including First Reasoning Model
Microsoft AI introduced a full-stack family of seven models trained from scratch with "zero distillation," led by the MAI-Thinking-1 reasoning model it claims rivals Claude Opus on coding benchmarks.
Key topics and reactions
Anthropic Warns AI Is Making Cyberattacks More Autonomous and Harder to Assess
Anthropic published an analysis of 832 accounts it banned for misuse of Claude between March 2025 and March 2026, mapping their activity to the industry-standard MITRE ATT&CK framework. The company warned that AI is making attackers more dangerous and attacks more autonomous, undermining conventional risk-assessment methods. Findings also contributed in part to Verizon's 2026 Data Breach Investigations Report.
Malware creation was the most common use at 560 accounts (67.3%), while lateral movement appeared in 54 accounts (6.5%). The share of cases rated medium-risk or higher rose from 33% in the first six months to 56% in the second. Phishing for initial access fell 8.6%, while post-intrusion activity rose, and nearly 80% of banned accounts used the agentic Claude Code tool.
Anthropic noted the correlation between attacker skill and number of techniques used is weakening — low-skill actors averaged 16 techniques versus about 20 for high-skill ones. The company introduced a new risk-scoring metric, ARiES, and said MITRE ATT&CK does not yet fully capture autonomous AI orchestration. Its Project Glasswing threat-intelligence effort has expanded to more than 150 organizations across over 15 countries.
Microsoft Unveils Seven In-House MAI Models at Build 2026, Including First Reasoning Model
At Build 2026, Microsoft announced its MAI model family spanning reasoning, coding, image, transcription and voice. The company says all seven were trained from scratch on cleanly licensed data without distilling from third-party models. The flagship MAI-Thinking-1 is a MoE model with 35B active parameters (about 1T total) and a 256K context window, reportedly scoring 52.8% on SWE-Bench Pro — comparable to Claude Opus 4.6 — and 97% on AIME 2025.
The lineup also includes MAI-Code-1-Flash (5B active) for agentic coding, claiming 71.6 on SWE-Bench Verified and up to 60% fewer tokens; MAI-Image-2.5 and 2.5-Flash; the 43-language MAI-Transcribe-1.5; and the emotion-controllable MAI-Voice-2 models. MAI-Image-2.5 has entered the top tier of image leaderboards, placing Microsoft among the top image-generation labs.
The models began rolling out the same day across Microsoft Foundry (some in private preview), GitHub Copilot, VS Code, PowerPoint and OneDrive, with availability via OpenRouter, Fireworks AI and Baseten. Microsoft AI, led by Mustafa Suleyman, framed the launch as the first step toward an automated "hill-climbing machine."
OpenAI Expands Life-Sciences Model GPT-Rosalind With GPT-5.5 Agentic Capabilities
OpenAI announced new features for GPT-Rosalind, its domain-specific model for life-sciences research first introduced on April 16, 2026. The update integrates GPT-5.5's autonomous coding and tool use to boost frontier reasoning for drug discovery, analysis, design and experimental workflows. The model is named after structural biologist Rosalind Franklin.
Positioned as an orchestration and reasoning layer distinct from structure-focused systems like AlphaFold, GPT-Rosalind supports multi-step workflows across literature, databases and tools. It remains a research preview limited to vetted U.S. enterprise customers, accessible via ChatGPT Enterprise, Codex and the API for internal research use only.
Separately, OpenAI promoted its Codex offering with a new "Time to Fly" branding campaign — a roughly 90-second hero film by agency Alto — and teased new additions to its developer Showcase Gallery, including a Codex-built space puzzle game called "Time to Fly."
Google Releases Open Gemma 4 12B for On-Device Multimodal Use
Google released Gemma 4 12B, an integrated multimodal model with an encoder-free design licensed under Apache 2.0. The company highlighted 100% on-device operation on laptops and edge devices, joining a growing wave of open-weight releases pushing toward local AI.
The launch underscores momentum in on-device LLMs, with vendors increasingly redesigning the PC around local agentic workflows. Gemma 4's laptop execution sits alongside other open releases driving the trend.
Google also continued promoting its native multimodal video model Gemini Omni through Google Flow, recommending a creative trend that applies surprising twists to real-world footage. Gemini Omni, unveiled at Google I/O 2026, replaces the earlier Veo line and integrates Gemini's reasoning for any-to-any generation and editing.
Category highlights
Microsoft to GA Work IQ APIs on June 16
Microsoft said its Microsoft 365 "Work IQ" intelligence layer will reach general availability on June 16, letting external apps and agents tap workplace context. The APIs span four domains — Chat, Context, Tools and Workspaces — and support A2A, MCP and REST, giving Copilot and custom agents access to work context, intent and organizational signals rather than raw data.
Anthropic Renames Claude Code Trigger to "ultracode"
Anthropic changed the trigger word for Claude Code's Dynamic Workflows from "workflow" to "ultracode" after the common term accidentally spawned dozens to hundreds of parallel agents. One user reported 130 Opus 4.8 agents launching from a late-night "workflow" prompt. Users can still trigger workflows by explicitly saying "use a workflow for this." Anthropic also published guidance reporting it automates 95% of internal analytics queries with Claude.
NVIDIA Open-Releases Omnimodal World Model Cosmos 3
NVIDIA released Cosmos 3, an omnimodal world model handling text, image, video, audio and action in a single model, aimed at accelerating physical AI for robotics and autonomous driving. The company also touted new skills for physical AI agents that it says will change how robotics development is done.
Gopuff and xAI Launch Grok-Powered Shopping Assistant "Go"
Gopuff and xAI launched "Go," an AI personal shopping assistant powered by Grok text, audio and image models. Using purchase history, time, location and real-time X signals, Go builds carts and takes voice or text orders, with delivery in as little as 15 minutes from Gopuff's 400-plus micro-fulfillment centers.
MiniMax M3 Draws Agentic-Workflow Users With 1M-Token Context
MiniMax M3's 1M-token context is winning fans for agentic operation, with SEO users describing it as "a jetpack for SEO" and running it as an overnight worker via Hermes Agent. Praise centers on a free, capable AI worker and huge context, though users flagged slow Ollama backends, quantization performance concerns and time gaps versus Claude Opus 4.7 on some agentic browsing tasks.
Key trends
OpenClaw Ships Windows Support, Draws 1,300+ at GitHub HQ
OpenClaw 2026.6.1 added Windows node hosting, a Skill Workshop and MiniMax M3 support, reportedly setting record npm downloads. The open-source personal AI agent also held "After Hours @ GitHub," a curated gathering at GitHub HQ during Build 2026 with a 1,300-plus waitlist and Twitch/Discord livestreams.
Video Tooling: Higgsfield, JoyAI-Echo, Volcengine, PixVerse
Video generation saw broad activity: Higgsfield released five AI plugins for Premiere Pro and After Effects enabling in-timeline generation, editing and background removal; JoyAI-Echo published an open code-and-weights model for 5-minute multi-shot videos with character and voice consistency; Volcengine launched Seedance 2.0-powered "Vibe Creating"; and PixVerse opened its CPP 2.0 creator program.