Open-source tsunami: 2.8T and 975B parameters

In 48 hours, two giant open-weight models reshaped the AI landscape: China's Kimi K3 (2.8T parameters, Fable 5 level) and America's Inkling from Mira Murati's Thinking Machines (975B). Plus the EU introduces mandatory content labeling in 17 days, and OpenAI automates red teaming. This week is a turning point.

Intro · Alex

This was the week when open models stopped being 'almost as good' — they became a real alternative. We have Kimi K3 from China, which according to its creators beats Fable 5 and GPT-5.6 Sol in coding, and costs $3 per million tokens instead of $10. We have Inkling from Mira Murati's team — an open-weight model with Apache 2.0 license that you can download and integrate into your infrastructure. In this edition I've traced five breakthrough moments: Chinese open-source at a new level, the first Thinking Machines model, OpenAI's automated bug hunter, the upcoming EU AI labeling mandate, and the first agent vulnerability on CISA's must-patch list. Plus two gems from the cyber world — prompt injection as defenders' weapon and an AI-generated script in a real attack.

What's worth knowing

Kimi K3 — the largest open-source model ever (2.8 trillion parameters)

China's Moonshot AI released Kimi K3 — an open-weight model with 2.8 trillion parameters that its creators say rivals Anthropic Fable 5 and outperforms GPT-5.6 Sol in coding benchmarks. API pricing is $3 per million tokens versus $5 at OpenAI and $10 at Anthropic. Weights will be free on GitHub within days. This is a blow to the business model of closed providers.

Inkling — the first Thinking Machines Lab model (Mira Murati) with 975 billion parameters

Former OpenAI CTO Mira Murati's company enters the game with Inkling — a multimodal MoE model (975B parameters, 41B active), 1M token context, and Apache 2.0 license. Weights are on Hugging Face, an NVFP4 quantized version is also available. Thinking Machines targets companies that want to own a model, not rent one.

GPT-Red — OpenAI automates red teaming with self-play

OpenAI previewed GPT-Red — an autonomous system that uses self-play to discover prompt injection and other security flaws in its own models. The system was used to secure GPT-5.6 Sol before launch. This could set a new security standard in the industry and accelerate regulation.

EU: 17 days to mandatory AI content labeling — Article 50 AI Act takes effect

On August 2, 2026, Article 50 of the EU AI Act comes into force, mandating machine-readable watermarking of all AI-generated content — audio, video, images, and text. The European Commission has already published the Code of Practice. Every company operating in the EU must adapt its pipelines under threat of sanctions.

Langflow on CISA list — first agent framework on must-patch list

CISA added CVE-2026-55255 for the Langflow framework to the Known Exploited Vulnerabilities list — the first time an AI agent-building framework made the federal list. The IDOR flaw allows authenticated users to run other users' flows. If you use Langflow, patching is mandatory — the vulnerability is actively exploited.

From the tech world

Prompt injection as defenders' weapon — Tracebit blocks AI agents

Researchers at Tracebit discovered that hiding prompt injection chains alongside passwords in AWS reduces AI agent attack effectiveness from 57% to 5%. Defenders have turned attackers' weapons against them — a simple, cheap defense deployable immediately.

AI-generated PowerShell in a real Active Directory attack

Huntress documented an attack where the perpetrator used a vibe-coded PowerShell script to map AD — the script had comments, and reported results to HTML. One of the first documented cases of AI being used to generate attack code in a production environment.

Tip of the day

Your own watermarked pipeline before the EU deadline

In 17 days, Article 50 of the AI Act mandates labeling of AI-generated content. Instead of waiting for ready SDKs from providers, build a simple watermarking wrapper around the APIs you already use. Just add a middleware proxy layer that intercepts the model's response and adds a watermark from a library like Stable Signature or TrustMark — an implementation of just a few dozen lines of code. Practical step: pull a list of AI endpoints in your organization (OpenAI API, self-hosted models, no-code tools). For each one, define whether it generates content subject to the mandate (text, image, audio, video). Based on that, deploy a proxy layer with watermarking. Start with one pipeline — text — and test today, not in 16 days.