AI News digest — April 12, 2026

Top Stories

Anthropic Bans OpenClaw Creator From Claude Access

Anthropic temporarily banned the creator of OpenClaw — an open-source project that reverse-engineered Claude’s interface — from accessing Claude’s API and web interface. The move sparked debate across the AI community about the boundaries of reverse-engineering AI services and the tension between open-source ideals and platform terms of service. Anthropic cited violations of their acceptable use policy, while the OpenClaw project maintainers argued their work fell under fair research use.

ChatGPT Launches $100/Month Pro Plan

OpenAI officially launched ChatGPT Pro, a new $100/month subscription tier positioned above the existing $20/month Plus plan. The Pro tier offers higher rate limits, priority access to new model releases, and extended context windows. The launch comes as OpenAI seeks to diversify revenue beyond enterprise deals and position itself for a potential IPO. The pricing puts it in direct competition with Claude’s premium tiers and Google’s Gemini Advanced.

Meta Muse Spark: New Frontier Model with 16 Built-In Tools

Meta released Muse Spark, a new AI model available on meta.ai featuring an impressive 16 built-in tools including a code interpreter, web search, and visual grounding capabilities. Available at meta.ai, the model is competitive with frontier models from OpenAI and Anthropic on standard benchmarks. Muse Spark is not released as open weights, breaking with Meta’s traditional open-source approach with Llama. Simon Willison provided an in-depth analysis of the model’s tool architecture and capabilities, noting the sophisticated tool-chaining behavior and the code interpreter’s ability to execute Python in a sandboxed environment.

Security & Safety

Anthropic’s Project Glasswing Finds Vulnerabilities in Every Major OS

Anthropic disclosed results from Project Glasswing (also known internally as Claude Mythos), a restricted security research program that gave Claude deep system-level access to find vulnerabilities. Claude found previously unknown vulnerabilities in every major operating system tested, including Windows, macOS, and Linux kernels. The program is restricted to vetted security researchers under strict responsible disclosure protocols. The results show what AI-assisted offensive security research can uncover, and why these capabilities need controlled access.

Developer Tools & Platforms

Claude Cowork Gets IT Admin Deployment Tools

Anthropic rolled out new IT administrator deployment tools for Claude Cowork, its enterprise collaboration product. The new features include centralized user provisioning via SCIM, granular permission policies, audit logging, and SSO integration improvements. IT teams can now manage Claude Cowork deployments at scale through a dedicated admin console.

GitHub Copilot Updates: Retiring Opus 4.6 Fast, Pausing Pro Trials

GitHub announced several changes to the Copilot ecosystem:

Retiring Opus 4.6 Fast from Copilot Pro+ — the fast variant of Anthropic’s Opus 4.6 model is being removed from the model selector in favor of newer, more capable options
Pausing new Copilot Pro trials — GitHub is temporarily halting new sign-ups for Copilot Pro trials while they restructure the offering
New /fleet multi-agent feature — a preview of multi-agent workflows where multiple Copilot agents can collaborate on complex tasks in parallel

Google AI Edge Eloquent: Free Offline AI Dictation

Google released AI Edge Eloquent, a free offline AI dictation app for iOS built with on-device ML. The app provides real-time speech-to-text transcription without requiring an internet connection, using Google’s latest small language models optimized for edge deployment. Google has been pushing to bring AI to mobile devices without cloud dependencies, and Eloquent fits that strategy.

Microsoft Removing Copilot Buttons from Windows 11 Apps

Microsoft began removing Copilot integration buttons from several built-in Windows 11 applications, reversing its strategy of embedding AI throughout the OS. Users had complained about cluttered interfaces. Copilot remains available as a standalone experience but will no longer be embedded in every system app.

Open Source & Models

Z.ai GLM-5.1: 754B Parameter Model Under MIT License

Chinese AI lab Z.ai released GLM-5.1, a 754-billion parameter model under the permissive MIT license. The model is competitive with proprietary frontier models on reasoning and coding benchmarks. The MIT license allows unrestricted commercial use, making it one of the most openly licensed large-scale models available.

HuggingFace: Safetensors Joins PyTorch Foundation

HuggingFace announced that Safetensors, their secure tensor serialization format, is joining the PyTorch Foundation as an incubating project. Safetensors provides a faster, safer alternative to Python’s pickle for saving and loading model weights — eliminating the arbitrary code execution vulnerability that plagued pickle-based formats. The move to the PyTorch Foundation signals the format is becoming a standard for ML model distribution.

SQLite 3.53.0 Released

The SQLite team released version 3.53.0, bringing performance improvements for large aggregate queries, better JSON path expression handling, and improved WAL mode concurrency. SQLite continues to be the most deployed database engine in the world and remains a critical piece of infrastructure for AI tooling and edge applications.

Interesting Projects

Poke: AI Agents as Easy as Texting

Poke is a new project that aims to make interacting with AI agents as simple as sending a text message. The platform provides a messaging-style interface for orchestrating multi-step AI workflows, making it easier for non-technical users to run AI agents for everyday tasks.

Astropad Workbench: Remote Desktop for AI Agents

Astropad launched Workbench, a tool that provides remote desktop access designed specifically for AI agents. This allows AI systems to interact with graphical desktop environments — clicking buttons, reading screens, and navigating applications that don’t have APIs. GUI-only software is a large chunk of what agents can’t reach through APIs, and Astropad’s approach fills that gap.

Legal & Ethics

Stalking Victim Sues OpenAI Over ChatGPT

A stalking victim filed a lawsuit against OpenAI, alleging that ChatGPT was used to facilitate harassment by generating personal information and crafting threatening messages. The case could set precedents for how much responsibility AI platforms bear when their tools are misused.

Sam Altman Responds After Molotov Cocktail Attack on His Home

Sam Altman publicly responded to a New Yorker article following a Molotov cocktail attack on his home. As AI’s societal impact grows, so do the personal risks for industry leaders.

This digest is curated from TechCrunch, The Verge, Simon Willison’s blog, HuggingFace Blog, and GitHub Blog.