Published April 10, 2026

#2: Mythos Too Dangerous to Release, GLM-5.1 Beats Claude, Microsoft Agent Framework Hits 1.0

This Week's Highlights

Anthropic Locked Down Its Most Powerful Model

Claude Mythos Preview found thousands of zero-days across major operating systems and browsers — it can not just identify vulnerabilities but write exploits and chain them together. Anthropic is not releasing the model publicly: only 9 companies (AWS, Apple, Google, Microsoft, Cisco, JPMorgan, CrowdStrike, Broadcom, Nvidia) got access for defensive security through Project Glasswing. The company committed $100M in API credits and $4M in donations to open-source security.

What this means: The first time Anthropic officially declared a frontier model too dangerous to release publicly. The practical consequence: the era when AI finds vulnerabilities faster than humans patch them — has arrived. Update dependencies more aggressively, pentest without LLMs in 2026 is an anachronism.

Anthropic

GLM-5.1 — An Open Model Beats Claude and GPT on SWE-Bench Pro

Z.AI released GLM-5.1 — 754B parameters under MIT license. 58.4% on SWE-Bench Pro vs Opus 4.6 (57.3%), GPT-5.4 (57.7%), Gemini 3.1 Pro (54.2%). The key feature — it does not degrade on long tasks: sustains 8 hours of autonomous execution and 6000+ tool calls in a single session, optimizing solutions over hundreds of iterations.

What this means: For the first time, an open model outperformed frontier proprietary ones on real SWE tasks. If you are hitting API limits or do not want to send code out — run it on your own hardware. Weights on Hugging Face, MIT license (commercial use unrestricted). For long-horizon agent runs it may beat Claude.

Venturebeat

Microsoft Agent Framework 1.0 — Semantic Kernel and AutoGen Merged Into One SDK

On April 3 Microsoft shipped Agent Framework 1.0 for .NET and Python. Production-ready: stable APIs, long-term support, built-in MCP and A2A (agent-to-agent) support. Connectors for Azure OpenAI, Anthropic, OpenAI, Amazon Bedrock, Google Gemini, and Ollama. For migration from Semantic Kernel or AutoGen there is an automated assistant.

What this means: If you are on .NET building agents — this is now the default, there was no real option before. For Python — an enterprise alternative to LangChain and CrewAI. Built-in A2A means your agents can talk to others regardless of framework.

Microsoft

New Tools

n8n AI Workflow Builder (beta)

Describe automation in plain English — n8n generates a starting workflow. Plus they removed active workflow limits across all plans: now you only pay per execution.

N8n

MCP v2.1 — Server Cards

New discovery standard for MCP servers via `.well-known` URL. Browsers, crawlers, and registries can learn a server's capabilities without connecting. The protocol crossed 97M installs in March.

Modelcontextprotocol

AMD PACE

Open-source optimization framework for LLM inference on 5th-gen EPYC CPUs. For those self-hosting open-source models without GPUs — a real way to cut latency.

Amd

Recipe of the Week

Multi-Agent Research Team with CrewAI

CrewAI is a framework where multiple AI agents work as a real team: each with a dedicated role, tools, and area of responsibility. We break down how to build a research crew with a coordinator and specialists — from role design to debugging your first real scenario.

Useful Link

Assessing Claude Mythos Preview's Cybersecurity Capabilities (Anthropic red team)

Technical report on how Mythos found zero-days, which scenarios were tested, and where the model failed. Mandatory reading for security engineers, a sobering look at current model capabilities for everyone else.

Stat of the Week

Model Context Protocol crossed 97M installs in March 2026 — every major AI provider now ships MCP-compatible tooling. In a year it went from an Anthropic experiment to the default way agents connect to tools.

All Issues