AI News digest — April 11, 2026
Happy Friday, devs. New models, pricing shakeups, open-source milestones, and a security research tool that found bugs in every major OS and browser. Here’s the rundown.
Anthropic Debuts “Claude Mythos” — A Model Built for Breaking Things
Anthropic launched Claude Mythos, a variant built exclusively for vulnerability research. Distributed through Project Glasswing, Mythos isn’t public (restricted to vetted security researchers).
In testing, Mythos found exploitable vulnerabilities in every major operating system and browser (Windows, macOS, Linux, Chrome, Firefox, and Safari). That’s the first high-profile demonstration of a frontier LLM purpose-built for offensive security work.
Why it matters: If you build security tooling, automated vulnerability discovery at this level changes the economics of both attack and defense.
Thomas Ptacek (Matasano / Latacora) published a piece called “Vulnerability Research Is Cooked,” arguing that AI is fundamentally changing the security research landscape. Read it if you work in offensive security.
Links:
Meta Launches Muse Spark — And It’s Not Open Source
Meta released Muse Spark, a creative AI model for image generation and design. In a break from Meta’s open-source tradition (LLaMA, etc.), Muse Spark is:
- Not open-weight — proprietary model
- Hosted only — no local inference
- Accessible via private API — waitlisted access
The Meta AI companion app is already #5 on the App Store, suggesting strong consumer adoption. But the developer community is split. Many see this as Meta abandoning its open-source identity for competitive positioning against OpenAI and Google.
Meanwhile, GLM-5.1 from Z.ai (formerly Zhipu) dropped with 754 billion parameters under an MIT license, with solid SVG and code generation capabilities. If you want an open frontier model to benchmark, that’s the one.
Why it matters: The gap between “open by default” (Meta historically) and “closed for competitive reasons” is widening. GLM-5.1 under MIT is rare at this scale.
ChatGPT Pro Plan Launches at $100/Month
OpenAI launched the ChatGPT Pro plan at $100/month, well above the $20/month Plus tier. Pro targets power users and professionals who need higher usage limits and priority access during peak times.
The pricing tiers now stack up as:
- Free — basic access
- Plus ($20/mo) — GPT-4 level access
- Pro ($100/mo) — maximum compute, priority queue
Also on the OpenAI front: the company published a safety blueprint for addressing child exploitation in AI-generated content, outlining detection and prevention strategies.
Links:
GitHub Copilot Gets Smarter — And More Transparent
Two updates from GitHub:
“Rubber Duck” Second Opinion Feature
Copilot now has a “Rubber Duck” mode that routes complex queries to multiple model families simultaneously and synthesizes a second opinion. Cross-model verification instead of relying on a single model’s output.
/fleet for Parallel Agents
The new /fleet command spawns multiple Copilot agents in parallel, each working on a different aspect of your codebase. Handy for large refactors or multi-file changes.
⚠️ Data Policy Change — Effective April 24
GitHub is changing its data policy: interaction data from Free, Pro, and Pro+ Copilot users will be used for model training unless users explicitly opt out. This takes effect April 24, 2026. If you’re on any of these tiers and care about your code/interaction privacy, opt out in your settings before the deadline.
Why it matters: Cross-model verification catches errors that single-model suggestions miss. But check your data privacy settings.
Hugging Face: Safetensors Joins PyTorch Foundation + TRL v1.0
Two milestones from the HF ecosystem:
Safetensors → PyTorch Foundation
Safetensors, Hugging Face’s secure tensor serialization format, is joining the PyTorch Foundation as an independent project. Safetensors has become the standard for safe model weight distribution, since it avoids the arbitrary code execution risks of Python’s pickle format.
If you distribute models, you should already be using safetensors. This move ensures long-term governance and neutrality.
TRL v1.0 Released
TRL (Transformer Reinforcement Learning) hit v1.0, giving a stable, production-ready library for fine-tuning language models with RL techniques (PPO, DPO, KTO, etc.). Key features:
- Stable API surface
- Improved memory efficiency for large models
- Better integration with PEFT and LoRA adapters
pip install trl==1.0.0Why it matters: TRL v1.0 means you can build production RLHF/DPO pipelines without worrying about breaking API changes. Combined with safetensors joining PyTorch Foundation, the open-source ML infrastructure is maturing quickly.
Google: Gemma 4 On-Device + AI Edge Gallery
Google released Gemma 4, pitched as “frontier multimodal intelligence on device.” The model runs locally on consumer hardware with competitive performance to larger cloud-hosted models.
Alongside it, the Google AI Edge Gallery app lets developers browse, test, and deploy on-device AI models across Android and iOS.
Google also launched AI Edge Eloquent, a free offline AI dictation app that runs entirely on-device with no cloud dependency.
Links:
Enterprise & Infrastructure
Anthropic’s $30B Run-Rate + Google/Broadcom Deal
Anthropic is reportedly hitting a $30 billion annual run-rate. Separately, Google and Broadcom are teaming up on a major infrastructure deal to expand AI compute capacity.
Claude Cowork — IT Admin Tools
Anthropic launched Claude Cowork, a suite of IT admin tools for deploying Claude across organizations. Features include:
- Company-wide policy controls
- Usage analytics and audit logs
- Zoom transcript to action items pipeline (automatic meeting summarization with task extraction)
Microsoft Copilot Buttons Removed from Windows 11
Microsoft is removing Copilot buttons from Windows 11 built-in apps. The company says it’s “streamlining the experience,” but it signals a recalibration of how aggressively they push AI into the OS.
Microsoft Open-Sources AI Agent Security Toolkit
Microsoft released an open-source toolkit for runtime AI agent security, with middleware and patterns for securing autonomous AI agents in production. Includes guardrails, audit logging, and permission scoping.
Snowflake → AI Platform
Snowflake is accelerating its transition from data warehouse to full AI platform, with new capabilities for model training, inference, and agent deployment within the Snowflake ecosystem.
Atlassian: Visual AI Tools + Third-Party Agents in Confluence
Atlassian is adding visual AI tools (diagram generation, wireframing) and support for third-party AI agents directly within Confluence.
Quick Hits
| Story | TL;DR |
|---|---|
| Tubi | First streaming service with a native ChatGPT app for content discovery |
| Florida AG vs OpenAI | Investigating whether ChatGPT played a role in a shooting incident |
| Bezos’ Project Prometheus | Bezos’ AI lab poached an xAI cofounder from OpenAI — the talent wars continue |
| Meta losing open-source identity | Industry observers note Meta’s shift from “open by default” to competitive model gating |
Today’s Takeaway
Two directions at once. On one side: proprietary, hosted, controlled (Muse Spark, ChatGPT Pro). On the other: open, modular, community-governed (GLM-5.1 MIT, Safetensors joining PyTorch Foundation, TRL v1.0). Which ecosystem you build on matters more now.
Claude Mythos finding bugs in every major OS raises the stakes for anyone building infrastructure. And GitHub’s data policy change is a reminder to audit where your code and interactions end up.
Thoughts on any of these? Drop a comment or find me on the usual channels.