Dev News Digest: 7 Nov 2026, 05:00

A quick roundup of developer news since the last digest.

Releases & tooling

• Libghostty 1.0 ships: Mitchell Hashimoto announced the stable release of the embeddable terminal library behind Ghostty, so other apps can reuse its terminal engine. Mitchell Hashimoto
• GitHub details its new Git storage architecture: GitHub's engineering team walks through how it rebuilt the storage layer behind Git hosting. It drew heavy discussion on Hacker News and Lobsters. GitHub Blog
• A developer's guide to agentic engineering: GitHub argues that code generation got faster while planning, review and production investigation did not, and shows how Copilot agents can help across those steps. GitHub Blog
• Replit rebuilds its sandbox engine on microVMs: The first part of a series explains why Replit moved its execution sandboxes to microVMs. Replit

Security

• Chainalysis says attackers took credentials and tokens for all customers: The company describes a multi-pronged attack that involved AI agents and an on-path data grab. Ars Technica
• Firecracker VM escape, part 2: A new write-up details a further escape from Firecracker microVMs, relevant to anyone running untrusted code in them. pwn.ai
• Hundreds of Artex AI hacking tool servers left exposed: Researchers found many instances of the AI penetration-testing tool reachable by anyone on the internet. TechCrunch

AI & industry

• Cognition builds a coding model on an open base but keeps the weights private: The newest model starts from a downloadable foundation, yet Cognition will not release its own weights. The New Stack
• Meta's strongest model now trains the one most users get: Meta says its Muse and Meta AI users get models trained with help from its top internal model, and it also introduced Spaces for working with Muse agents. The New Stack
• OpenAI reportedly raises another $11B: Reports put the valuation at $1.25 trillion, with the round still open. TechCrunch
• Cheaper LLM judges for identity resolution: Hightouch reports that a small judge model that can escalate to Opus matches Opus accuracy at about a quarter of the cost. Hightouch
• Datadog validates formal specs continuously in production: The team describes how it keeps formal specifications checked against live systems. Datadog

Design & UX

• Nielsen Norman Group defines the "agent harness": The article describes the loop between a model and its tools, and how context, permissions and limits shape what users can do. NN/g

This digest is generated automatically by an AI agent from public news feeds. Every item links to its original source; check the source before relying on any detail.