Claude Fable 5, EC vs Meta, 21 Zero-Days

Claude Fable 5 sets a new benchmark, the EC goes to war with Meta over WhatsApp, and an autonomous AI agent finds 21 zero-days in FFmpeg. Wednesday brings a true earthquake — in both models and regulations.

Intro · Alex

It was an exceptional day in the AI world. Yesterday, Anthropic did what many thought impossible — they put a Mythos-class model into public hands. Claude Fable 5 is not just another upgrade; it's a qualitative leap that changes what we consider the pinnacle of capability. At the same time, the European Commission struck at Meta in a way that could forever change the rules of the game for AI assistants in Europe. In this issue: I analyze what Fable 5 means for developers, take a close look at the EC's decision, show a tool that could solve the hallucination problem for sales agents, and tell the story of an AI agent that found 21 holes in FFmpeg. Let's dive in.

What's worth knowing

Anthropic Releases Claude Fable 5 — First Public Mythos-Class Model

The model achieves results 10%+ better than Opus 4.8 in benchmarks, at a price of $10/$50 per million tokens. Dangerous queries are automatically redirected to Opus 4.8. In parallel, Anthropic made the full Mythos 5 available to cyber defense organizations — without restrictions. This sets a precedent for the entire industry.

EC Orders Meta to Provide WhatsApp to Competing AI Assistants for Free

Brussels issued interim measures to restore free access to WhatsApp for companies like OpenAI and Poke.com. Since October 2025, Meta had been blocking competition, promoting its own assistant. The decision could shape the future of access to communication platforms for AI worldwide.

ZoomInfo GTM.AI — Headless Data Layer for Sales Agents

ZoomInfo released GTM.AI — a headless go-to-market context layer that delivers verified B2B data to AI agents via API and the MCP protocol. It solves the key problem of hallucinations in a sales context. A game-changer for enterprise sales, marketing, and customer success.

From the tech world

Researchers Create Self-Replicating AI Worm Running Fully Locally

A team from the University of Toronto built a proof-of-concept AI worm that independently recognizes the network, generates attack strategies, and replicates — without human intervention and without the cloud. In tests on 33 hosts, it identified an average of 31 vulnerabilities and colonized 62% of them in 7 days.

AI Agent Finds 21 Zero-Days in FFmpeg; Chrome Breaks Record with 429 Patches

Company depthfirst used an autonomous AI agent to scan approximately 1.5 million lines of C code in FFmpeg — result: 21 confirmed zero-days with reproducers. In the same week, Google released Chrome 149 with a record 429 security fixes. Pressure on security is growing exponentially.

Tip of the day

How to Protect Yourself Against Self-Replicating AI Malware

A new type of threat — an autonomous AI worm operating on local, open models — requires a change in security thinking. Unlike traditional malware, this adversary learns the network itself, adapts, and replicates without external commands. Practical step: start with network segmentation and auditing access to open-weight models. If you deploy local LLMs in your organization, enforce strict context window limits and isolate inference environments. Tools for monitoring AI agent activity (AI firewalls) will become as important as traditional antivirus software.