At the Build conference on June 2, Microsoft unveiled seven new in-house AI models, including the flagship reasoning model MAI-Thinking-1 (35B active parameters / 1T total parameters MoE architecture), coding model MAI-Code-1-Flash, image model MAI-Image-2.5, transcription model MAI-Transcribe-1.5, and speech model MAI-Voice-2. MAI-Thinking-1 matches Claude Opus 4.6 on SWE-Bench Pro, achieves 97% accuracy on AIME 2025, and outperforms Sonnet 4.6 in blind user preference tests. All models are trained from scratch on clean commercially licensed data without third-party distillation. Microsoft also introduced Frontier Tuning technology, enabling up to 10x efficiency gains in applications like Excel. The company is collaborating with Mayo Clinic to develop medical AI models. The models are now available via platforms such as Foundry and OpenRouter.

On June 2, President Trump signed an executive order establishing a voluntary pre-release review framework for frontier AI models. The order allows AI companies to voluntarily submit their models to the government up to 30 days before release for cybersecurity assessment—significantly shortened from the previous 90-day draft requirement. It directs the Treasury, Defense, and other departments to establish an AI Cybersecurity Information Sharing Center to strengthen critical infrastructure defense, and instructs the Justice Department to prioritize prosecution of AI-enabled cybercrimes. The order explicitly avoids mandatory licensing. Google, Microsoft, and xAI have agreed to participate in testing. The discovery of numerous high-risk vulnerabilities by Anthropic's Mythos model was a key catalyst for this policy. This marks a shift in the Trump administration’s AI stance from laissez-faire to proactive oversight.

On June 2, OpenAI launched a major update to its Codex platform, introducing three core features: Sites enables users to quickly create interactive, hosted web applications using natural language; Annotations allow precise, localized editing of documents like spreadsheets; and six categories of role-specific plugins integrate 62 mainstream business applications (e.g., Snowflake, Figma, Salesforce) and 110 automation skills across data analysis, creative production, and sales. Codex currently has 5 million weekly active users, with non-developers making up about 20%—and adopting the tool at three times the rate of engineers. New features are available through existing subscriptions, starting at $20/month for Plus. This update signals Codex’s evolution from a programming tool to an enterprise knowledge work operating system.

At Computex, NVIDIA unveiled the open-source Nemotron 3 Ultra 550B model, featuring 550 billion parameters, LatentMoE architecture, Mamba-2 layers, and support for a 1 million token context window, achieving inference speeds over 300 tokens/second. The model scored 48 on the AI Index—the highest among U.S. open models—but trails China's Moonshot Kimi K2.6, which scored 54. NVIDIA also launched an Agent Toolkit (including NemoClaw, OpenShell, and CUDA-X skill libraries), already adopted by enterprises like Cadence, CrowdStrike, and Palantir. Model weights will be released on Hugging Face on June 4, with four deployment options. The strategy focuses not on building the smartest model, but on driving enterprise AI adoption on NVIDIA hardware through free models.

At Build 2026, Microsoft introduced Scout, an always-on personal AI assistant built on the open-source OpenClaw framework. Unlike Copilot, Scout runs continuously in the background, monitoring calendars, traffic, and communications to proactively suggest actions and manage tasks. Users can name and continuously train it to automate workflows, creating personalized loops. Scout includes a built-in compliance system that monitors behavior and generates audit logs to prevent AI misuse. Over 3,000 Microsoft employees are already using it internally. Scout will first launch as a desktop preview for U.S.-based Frontier customers, requiring a GitHub Copilot subscription. Microsoft also announced Project Solara, a new operating system platform for AI agent devices.

After encouraging employees to actively use AI coding tools (e.g., Claude Code and Cursor) and gamifying usage with leaderboards, Uber exhausted its entire annual AI budget in just four months, forcing it to impose a $1,500 monthly spending cap per employee. The company has established an internal dashboard to track usage, with exceptions allowed upon approval. CEO Andrew Macdonald previously questioned AI’s productivity impact, stating, "It’s hard to draw a line between AI usage and new consumer features." This incident reflects a broader industry challenge: surging enterprise AI spending with unclear ROI, which remains, for now, a theoretical phenomenon everyone hopes will eventually materialize.

GitHub has launched a new Copilot desktop application as a unified control center for managing multiple AI agents running in parallel. Key features include isolated git worktrees to ensure session separation, a Canvas bidirectional interface enabling human-agent collaboration where developers can directly edit agent plans and code, and cloud and local sandboxes for secure execution. Code review now includes a medium severity level and specialized skills like /security-review. Additionally, the Copilot SDK has reached general availability, supporting multiple languages including Node.js, Python, and Go, allowing developers to embed Copilot’s agent engine into their own applications. GitHub is currently facing infrastructure challenges due to a 1400% increase in AI-generated code volume.

HCompany has released the Holo3.1 family of computer-use agent models, supporting cross-platform operation on web, desktop, and mobile. The 35B-A3B model improved accuracy on AndroidWorld from 67% to 79.3%, added function calling protocol support, and achieved over 25% performance gains across frameworks compared to the previous generation. It introduces quantized checkpoints (FP8, Q4 GGUF, NVFP4), with NVFP4 delivering a 1.74x throughput improvement on DGX Spark and reducing end-to-end step time from 6.8 seconds to 3.3 seconds. Four model sizes—from 0.8B to 35B-A3B—are available, all supporting fully local deployment to ensure data privacy. Models are accessible via the Holo Models API and Hugging Face.

Security researcher RyotaK discovered a critical supply chain vulnerability (CVSS 7.8) in Claude Code’s GitHub Actions, caused by the checkWritePermissions function incorrectly trusting all participants ending in [bot], leading to privilege escalation. Attackers could create a malicious GitHub App, use installation tokens to trigger workflows via issues in target repositories, obtain OIDC tokens, and exchange them for write-access tokens—even injecting malicious code into Anthropic’s own claude-code-action repository. Other researchers have highlighted risks of prompt injection attacks against AI coding agents in CI environments, where secrets can be exfiltrated via GitHub PR titles and comments. Anthropic patched the issue in version v1.0.94 and awarded the researcher a $4,800 bounty.

Japan’s Minister for Financial Services, Katayama, announced late on June 2 that the Japanese government and several domestic banks have been granted access to Anthropic’s high-performance AI model, Claude Mythos. Mythos has drawn attention for its exceptional ability to discover system vulnerabilities, raising concerns about potential misuse leading to large-scale cyberattacks such as fund transfer failures. Currently, around 200 global organizations across more than 15 countries have access, with Japan now among them. On the same day, Anthropic announced Project Glasswing, expanding Mythos access to 150 new partner institutions in critical infrastructure sectors including power, water, and healthcare, having already identified over 10,000 high-risk security flaws since April.