Global AI Weekly - Issue 161 - Global AI Community

In this issue

Highlights

2 stories

reuters.com

OpenAI to triple workforce at Dublin European headquarters to 350

OpenAI will more than triple its Dublin headcount to 350, announcing leases for 8,000 square metres in the Silicon Docks and plans to hire 250 staff over two years. Recruiting will focus on engineering and support operations, expanding a European hub that already employs just over 100 people. For developers and architects, expect increased local talent availability, faster onshore support for EU deployments, and tougher competition for senior AI engineers. It reshapes Irish AI hiring.

github.blog

Grok 4.5 is now available in GitHub Copilot

GitHub Copilot now includes xAI's Grok 4.5, a reasoning model tuned for agentic coding and multi-step workflows. It supports text and image inputs, selectable low/medium/high reasoning effort, and a 500,000-token context window. In GitHub's internal tests, it excelled at parallel tool dispatch and terminal-based coding tasks. Grok 4.5 is rolling out to Pro, Pro+, Max, Business, and Enterprise across editors and CLI; admins must enable the model and expect usage-based billing implications, and monitor costs.

In this issue

Research

2 stories

github.com

MoonshotAI releases Kimi K3, a 2.8T open-weight model with 1M-token context

MoonshotAI released Kimi K3 and open-sourced its weights, a 2.8 trillion-parameter model with native multimodality and a 1M-token context window. The architecture pairs Kimi Delta Attention and Residuals with a LatentMoE activating 16 of 896 experts, yielding 2.5× scaling efficiency over Kimi K2. For long-horizon coding and agentic knowledge work, Kimi K3 uses MXFP4/MXFP8 quantization-aware training and a 401M-parameter MoonViT-V2 vision encoder. Open weights let engineers reproduce, fine-tune, and test safety at scale.

microsoft.com

Deep, evolving environments for computer-use agents

Microsoft Research released Echoverse, a pipeline and twelve training worlds for computer-use agents, and released four with code, data, and grounded verifiers. The worlds emphasize depth: faithful state, workflows, and SQL-rooted verifiers that grade outcomes, not screenshots. Trained on twelve, a 9B model rose from 36.5% to 67.1%, and reinforcement learning against the worlds further improved held-out performance. If you build agents that must click, write, or maintain state across screens, Echoverse is worth testing.

In this issue

Video

2 stories

youtube.com

Agents All the Way Down: How the VS Code Team Builds with What It Ships

Commits doubled. Issues closed tripled. Releases went from monthly to weekly — same team size. The VS Code team wired agents into the full product cycle: a PM turns social feedback into a working PR before the next meeting, an EM kicks off parallel agent sessions from his calendar each morning, and issue triage runs as an agent loop in GitHub Actions. Custom agents, parallel workflows, and the patterns you can steal.

youtube.com

Understanding AI Agent Hallucination in AI Systems

Learn more about AI Hallucinations here → https://ibm.biz/~UAoGwgNhw Confident AI answers are not always grounded in truth. Brianne Zavala explains what agent hallucination is and why it still happens. Learn how tools, data grounding, and design choices reduce hallucination risk. AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM → https://ibm.biz/~mOAD4HcZb AI was used in the creation of the transcript and metadata for this video. #aiagents #aihallucinations #agenticai #aitools #humanintheloop

In this issue

Articles

2 stories

microsoft.ai

Introducing MAI-Cyber-1-Flash inside MDASH

Microsoft released MAI-Cyber-1-Flash inside MDASH, a compact code-focused model integrated into a multi-agent vulnerability identification and remediation harness. The combined MDASH+MAI system hits 96% on CyberGym, 12 points above Mythos, and reduces model costs by about 50% compared with Microsoft's prior MDASH stack. MAI-Cyber-1-Flash is tuned to handle roughly 90% of routine scans while deferring the hardest 10% to larger models, and Microsoft adds Perception agents for continuous monitoring and patching, and tenant isolation capabilities.

techcrunch.com

Sam Altman is still making the case for parenting via ChatGPT

Sam Altman proposed parents could connect family calendars to ChatGPT Work and have it generate a morning 'podcast' for kids, a Twitter post that went viral. The reaction highlighted consumer skepticism, with Alex Hirsch's "What if you just talked to your children" outpacing Altman's post; it arrived as OpenAI recruited for family-focused product roles and defended itself in lawsuits alleging harm tied to ChatGPT. The episode amplifies UX, trust, and liability questions teams must solve.

In this issue

Upcoming Events

1 story

globalai.community

Upcoming Events Around the World

From local meetups to full-day conferences, the Global AI Community brings developers, architects, founders, and AI enthusiasts together to learn, build, and connect. Discover hands-on workshops, expert talks, community gatherings, and flagship events like AgentCon, AgentCamp, and Global AI Construct. Whether you are exploring AI for the first time or building production-ready agentic systems, you will find opportunities to sharpen your skills, meet like-minded people, and become part of a thriving global community.

In this issue

Code

3 stories

github.com

OpenAI open-sources Codex Security CLI and TypeScript SDK

OpenAI published codex-security, an open-source CLI and TypeScript SDK that finds, validates, and helps fix security vulnerabilities in codebases. The package, @openai/codex-security, supports Node.js 22/24/26, offers interactive and CI-friendly authentication (API key or ChatGPT sign-in), and exposes deep-scan modes, worker parallelism, and containerized bulk scans with resumable state. Scans compare before/after runs to track new, persisting, and resolved findings. For teams automating security, the repo is a practical starting point for Git-based, GPT-assisted code scanning.

devblogs.microsoft.com

Discover Agent Skills from MCP servers in .NET

Microsoft's Agent Framework for .NET lets agents discover and load Agent Skills directly from Model Context Protocol (MCP) servers using the Microsoft.Agents.AI.Mcp package. Skills can be served as SKILL.md resources or as archives that the framework downloads, validates, and extracts under controlled limits. Remote scripts are never executed, and archive size, file count, and uncompressed size are configurable to prevent resource abuse. For multi-agent deployments, this centralizes skill distribution and enables updates without redeploying agents.

towardsdatascience.com

How to Give an LLM Agent a Browser

In this issue

Podcast

1 story

open.spotify.com

The AI Policy Podcast

Listen to The AI Policy Podcast on Spotify. Join CSIS’s Aalok Mehta, Director of the Wadhwani AI Center, on a deep dive into the world of AI policy. Every two weeks, tune in for insightful discussions covering AI regulation, economic impacts, national security, geopolitics, and more. CSIS is a bipartisan think tank in Washington, D.C.