Issue #176 13 stories

Global AI Weekly

Claude Fable 5.2 Takes AI Engineering to Another Level

Published Tuesday, November 17, 2026

In this issue

Highlights

2 stories
Claude Fable 5.2 Takes AI Engineering to Another Level
anthropic.com

Claude Fable 5.2 Takes AI Engineering to Another Level

Anthropic has released Claude Fable 5.2 and Mythos 5.2, two versions of its most powerful model, built for demanding technical challenges. The company reports substantial improvements in coding, reasoning, science and robotics, alongside faster output and lower pricing. Early demonstrations include designing working electronic devices, migrating production software and coordinating thousands of agents on complex projects. Fable is broadly available with additional safeguards, while Mythos requires verified access for specialized cybersecurity and biological research.

Google's AI Just Mathematically Verified Signal's Security
deepmind.google

Google's AI Just Mathematically Verified Signal's Security

Google DeepMind has introduced AlphaProof Code, a system that uses AI agents and mathematical proofs to verify real software. In a striking demonstration, researchers proved key security properties of Signal's cryptographic implementation, down to machine code. The project generated millions of lines of machine-checked proofs in just four weeks. Human experts approved the security specifications, while AI agents handled much of the verification. This could change how critical software is audited, from messaging applications to financial infrastructure.

In this issue

Research

2 stories
AI Takes on GTA V, and Driving Is Still Hard
arxiv.org

AI Takes on GTA V, and Driving Is Still Hard

Researchers have turned Grand Theft Auto V into a testing ground for AI agents with GTAGym, a framework designed to evaluate perception, reasoning and real-time control. The benchmark includes 103 reconstructed game tasks and 47 additional image-based challenges. Agents interact through screenshots, keyboard inputs and mouse actions, with outcomes checked automatically. The strongest tested configuration reached 70.9% accuracy on one benchmark, but aerial navigation and evasive driving remain difficult. Apparently, even advanced AI struggles with Los Santos.

AI Agents Get a 4.1x Speed Boost Without a New Model
arxiv.org

AI Agents Get a 4.1x Speed Boost Without a New Model

As AI agents become more complex, serving their models efficiently is increasingly important. Researchers introduce Evolving Infra, a framework that dynamically reallocates GPU resources and manages cached context according to agent workloads. Rather than treating every request alike, the system adjusts how computing resources handle incoming prompts and generated tokens. In experiments involving coding, mathematics and computer-use agents, it reduced average end-to-end latency by up to 4.1 times compared with NVIDIA Dynamo, without changing the underlying models.

In this issue

Video

2 stories
Watch AI Hunt Real Firefox Security Bugs
youtube.com

Watch AI Hunt Real Firefox Security Bugs

Can AI find and fix vulnerabilities in software used by millions? In this discussion, Mozilla's Brian Grinstead and OpenAI's Clint Gibler explore how GPT-6 Cyber is being used to identify security weaknesses in Firefox. The conversation moves beyond benchmark scores to examine practical AI-assisted vulnerability discovery and remediation. It offers an interesting look at where specialized cybersecurity agents can help experienced engineers, and why reviewing their findings and fixes remains essential before shipping security-sensitive changes.

Run Claude Code Locally Without API Bills
youtube.com

Run Claude Code Locally Without API Bills

What if you could use Claude Code without paying for cloud-based model inference? Mervin Praison demonstrates how to connect the coding environment to Ollama and a locally running model. The tutorial covers configuration, tool compatibility and a practical workaround for command-naming issues that initially prevent tasks from completing. He then tests the setup against ten coding jobs. It's an interesting option for developers experimenting with private, local coding workflows and existing hardware.

In this issue

Articles

3 stories
Why everyone is suddenly rebuilding Git
Global AI Community

Why everyone is suddenly rebuilding Git

Anthropic replaced Git, OpenAI rebuilt its engine and is buying a code hosting startup, and Entire, Cursor and GitHub are all racing to host agent-written code. We look at why AI agents broke the scale Git was built for, and what it means for your team.

Your AI Agent Can Now Shop, but You Approve the Bill
openai.com

Your AI Agent Can Now Shop, but You Approve the Bill

OpenAI has introduced ChatGPT Wallet, allowing AI agents to complete purchases without accessing a user's actual card details. Developed with American Express, Mastercard and Visa, the system uses payment tokens and requires explicit approval, including a biometric security check, for each transaction. Initial availability is limited to eligible US ChatGPT Pro subscribers using dots. It's an important step toward agent-driven commerce, although merchants, payment providers and users still need to agree on how autonomous shopping should work.

In this issue

Upcoming Events

1 story
Microsoft Ignite 2026: The Next Wave of Enterprise AI
ignite.microsoft.com

Microsoft Ignite 2026: The Next Wave of Enterprise AI

Microsoft Ignite returns to San Francisco, bringing developers, IT professionals and business leaders together to explore what's next in AI, cloud computing and cybersecurity. This year's agenda puts significant attention on enterprise AI agents, Microsoft Foundry, infrastructure and secure deployment. Attendees can expect product announcements, technical sessions, live demonstrations and hands-on labs. With online access available, the event offers a chance to explore new capabilities and examine which developments are ready for real-world adoption.

In this issue

Code

2 stories
Cloudflare Just Open-Sourced a Browser Built for AI Agents
github.com

Cloudflare Just Open-Sourced a Browser Built for AI Agents

Cloudflare has released Kitesurf as an Apache 2.0 open-source project, giving developers access to its experimental browser engine designed specifically for AI agents. Unlike conventional browser automation, Kitesurf runs on Cloudflare Workers without requiring a full Chromium container. It supports HTML, CSS, JavaScript execution and the Chrome DevTools Protocol, allowing integration with tools such as Puppeteer. Developers can inspect the code, experiment locally and deploy their own instances, although compatibility remains a work in progress.

Write an Agent Skill Once, Use It Everywhere
github.blog

Write an Agent Skill Once, Use It Everywhere

Tired of explaining the same workflow to every coding assistant? GitHub's practical guide shows how to package reusable instructions as agent skills, allowing teams to share workflows across tools such as GitHub Copilot, Claude Code and Cursor. Using SKILL.md files, project-specific configuration and automated checks, developers can build portable, reviewable capabilities. The article includes examples for code verification and data pipelines, with an emphasis on clear responsibilities, reliable testing and human approval.

In this issue

Podcast

1 story
AI Agents That Run for Months? Here's What It Takes
open.spotify.com

AI Agents That Run for Months? Here's What It Takes

What does it take to keep an AI agent running for weeks or even months? Conor Bronsdon talks with Inngest CEO Tony Holdstock-Brown about the infrastructure behind long-running autonomous workflows. They discuss durable execution, checkpoints, retries, persistent context, subagents and the growing pressure that agents place on APIs. The conversation challenges the idea that better models alone will solve agent reliability, explaining why traditional distributed-systems engineering is becoming essential to practical AI applications.

Every Tuesday

Get the next issue in your inbox.

The AI links worth your time, curated by the community and always free.