AI Dev News Digest: December 19th, 2025

This week was another AI launch rush. OpenAI released GPT‑5.2‑Codex with security tools that found React bugs. They also opened the ChatGPT App Directory and hired a Shopify exec to make it an OS.

Google answered with Gemini 3 Flash worldwide and added TypeScript support for its Agent Development Kit. Apple quietly shared a mind blowing 2D‑to‑3D model. Anthropic pushed its Skills format into an open standard, building a partner ecosystem. The agent AI race is heating up.

Other news: MIT Technology Review said 65% of developers use coding tools every week, but they still don’t boost productivity as much as people think. Mozilla announced Firefox will be an “AI browser” and got backlash from its users. Vectorize launched an agent memory system that scored 90% on long‑term tests. Security experts warn that prompt injection and “vibe coding” could become major threats.

On Reddit, the Era of the Idea Guy meme is trending where someone with a four‑button keyboard asks Claude to build a Mars pet social network.

And so much more!

Foundation Models

• OpenAI releases GPT-5.2-Codex for agentic coding and cybersecurity. The model adds context compaction for million-token tasks, better refactoring, and notably stronger security capabilities. A researcher used it to find three React Server Components vulnerabilities. Available to paid ChatGPT users now, API coming soon. (OpenAI)

• Google launches Gemini 3 Flash globally. Replaces 2.5 Flash as the default in the Gemini app and Search AI Mode. Scores 33.7% on Humanity's Last Exam without tools and 81.2% on MMMU-Pro. Pricing is $0.50/$3.00 per million tokens. JetBrains, Cursor, and Figma already using it. (TechCrunch)

• OpenAI ships new ChatGPT Images with GPT Image 1.5. 4x faster image generation with better edit controls, consistent lighting, and facial likeness preservation. Available to all ChatGPT users and in the API. (OpenAI)

• Nvidia releases Nemotron 3 open models. Three sizes: Nano (30B), Super (100B), and Ultra (500B). Hybrid mixture-of-experts architecture with 1M token context. Designed for multi-agent systems. Already adopted by Perplexity, Palantir, and CrowdStrike. (Nvidia)

• xAI launches Grok Voice Agent API. Real-time voice for building voice agents at $0.05/min flat rate. Multiple voice personas, multilingual support, and sub-second latency. Same tech as Tesla vehicles. (xAI)

• Apple open-sources SHARP for instant 2D-to-3D conversion. Feed it a single photo, get a photorealistic 3D scene in under a second using Gaussian splatting. 25-34% better quality scores than previous approaches and 100+ fps rendering. Code and weights on GitHub. (Apple ML)

AI Coding Tools

• GitHub Copilot now supports Agent Skills. Create folders with instructions and scripts that Copilot loads automatically for specific tasks. Works across Copilot CLI, coding agent, and VS Code Insiders. Picks up Claude Code skills from .claude/skills automatically. (GitHub)

• Claude Opus 4.5 goes GA in GitHub Copilot. Available across GitHub.com, Mobile, VS Code, Visual Studio, JetBrains, Xcode, and Eclipse. Enterprise admins need to enable the new policy in Copilot settings. (GitHub)

• GPT-5.2 now generally available in GitHub Copilot. Access through the model picker in all supported IDEs. Part of Copilot's multi-model strategy alongside Claude and Gemini. (GitHub)

• GitHub Copilot adds Gemini 3 Flash in public preview. Available across Pro, Pro+, Business, and Enterprise tiers in VS Code, GitHub.com chat, and Mobile. Ideal for tasks where speed matters. Enterprise admins must enable the Gemini 3 Flash policy in settings. (GitHub)

• Claude Code adds Chrome browser control beta. Work with the Claude in Chrome extension to control your browser directly from Claude Code. Also adds config toggles for prompt suggestions and fixes for MCP loading. (Anthropic)

• OpenAI GPT-5.2-Codex hits state-of-the-art on SWE-Bench Pro. 56.4% accuracy and 64% on Terminal-Bench 2.0. Vision support for interpreting screenshots and UI mocks. Invite-only trusted access pilot for cybersecurity professionals. (OpenAI)

Agentic Infrastructure

• Anthropic launches Agent Skills as an open standard at agentskills.io. Skills are portable folders of instructions and scripts that work across Claude, Claude Code, Cursor, GitHub Copilot, Amp, Goose, and VS Code. Enterprise admins can now provision skills org-wide from admin settings. (Anthropic)

• Anthropic launches Skills Directory with partner integrations. Prebuilt skills from Atlassian (Jira backlogs, Confluence knowledge), Canva (campaigns, presentations), Figma (design-to-code), Vercel (deploy from chat), Cloudflare (one-shot agent deployment), and Zapier (workflow automation). Browse at claude.com/connectors. (Anthropic)

• OpenAI opens ChatGPT App Directory to developers. Third-party apps can now run inside ChatGPT with buttons, maps, sliders, and other UI elements. Built on Anthropic's MCP. Trigger apps with mentions or from the tools menu. (VentureBeat)

• Copilot code review preview features hit Enterprise Cloud. Agentic tool calling for full project context, integration with CodeQL and ESLint, and handoffs to Copilot coding agent for fixing issues. (GitHub)

• Nvidia acquires SchedMD for Slurm workload management. Slurm handles scheduling for HPC and AI training clusters. Will remain open-source and vendor-neutral. Critical infrastructure for foundation model training. (Nvidia)

• Google ships Agent Development Kit for TypeScript. JS/TS developers can now build AI agents code-first with modular components and testable architecture. Full support for Gemini 3 Pro and Flash, model-agnostic design, and MCP integration. (Google Developers)

• Vectorize releases Hindsight, an open-source agent memory system. First to break 90% accuracy on LongMemEval benchmark at 91.4%. Organizes memory into world knowledge, experiences, opinions, and observations. Developed with Virginia Tech and Washington Post researchers. (PR Newswire)

Industry News

• MIT Tech Review: The great AI hype correction of 2025. 95% of businesses found zero value in AI per one MIT study. Upwork research shows agents from OpenAI, Google, and Anthropic failed many straightforward workplace tasks. Reality setting in. (MIT Tech Review)

• MIT Tech Review: AI coding is everywhere but not everyone is convinced. 65% of developers now use AI coding tools weekly per Stack Overflow. Stanford study finds junior developer employment down 20% since 2022. The productivity vs. code quality debate continues. (MIT Tech Review)

• SD Times: 2025 was the year AI influenced all areas of software development. At least 15 companies launched AI coding tools this year. The downside: hallucinations and overwhelmed code reviewers. GitKraken calls it the "Builder's Era." (SD Times)

• Mozilla says Firefox will become an "AI browser," sparks major backlash. New CEO Anthony Enzor-DeMeo announced a 3-year pivot to AI features. Users are not pleased, calling it out of touch with Firefox's privacy-focused identity. AI features will be optional and easy to disable. (Windows Central)

• OpenAI hires ex-Shopify exec to turn ChatGPT into an OS. Glen Coates joins as Head of App Platform reporting to Nick Turley. His job: evolve ChatGPT from a chatbot into an operating system where apps live. Ties into the Jony Ive device expected in 2027. (Digital Trends)

• Promon warns prompt injection attacks will surge as AI moves to mobile. Security expert predicts prompt injection will become one of the fastest-growing mobile threats in 2026. Also forecasts 30% of new security exposures by 2027 may stem from "vibe coding." (Security Brief)

Agentic AI Foundation

• Linux Foundation announces the Agentic AI Foundation (AAIF). Anthropic donates MCP, Block donates goose, OpenAI donates AGENTS.md. Platinum members include AWS, Google, Microsoft, Bloomberg, and Cloudflare. MCP now has neutral governance. (Linux Foundation)

Weekend Read

• Agent Skills Getting Started Guide — Want to future-proof your AI workflows? Dive into this guide to learn how Agent Skills work, why they’re becoming the new standard, and how you can start building and using them today. Perfect for anyone looking to level up with the next generation of AI tools.

Weekend Watch

• Anthropic Skills Deep Dive — If you want to understand how Agent Skills actually work and why they matter, this walkthrough covers the full stack: creating skills, the open standard at agentskills.io, partner integrations, and enterprise deployment. Worth the time if you're building with any AI coding tool.