Dakarda AI Newsletter · 2 September 2026
Wednesday Edition
The same model, two worlds of security
Anthropic did something I didn't expect so soon: it released two models of the same architecture, but with diametrically different security measures — Fable 5.1 for everyone and Mythos 5.1 for the select few. In the background, Brussels uses the AI Act against the biggest labs for the first time.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
Welcome back after the break. 72 hours have passed, and the world of AI has shifted — not just by a few more percentage points in benchmarks, but by an entire distribution model. Anthropic showed that the same model core can reach the public in a restricted version and selected organizations in the Mythos variant with looser barriers. This isn't about "more power" — it's about who we want to give it to and how. In this edition, I'll walk you through Anthropic's new model pair, Brussels' first real enforcement of the AI Act, and OpenAI's upcoming Astra model — so capable that its own creator says it "requires additional safeguards." Plus, Atlas from World Labs — a model that understands 3D space — and a practical video editing tool called Genjutsu. We'll cover architecture, law, and one bug that scored a maximum CVSS 10.0.
What's worth knowing
Anthropic releases Claude Fable 5.1 and Mythos 5.1 — same core, two worlds of security
Anthropic has officially released two models: the publicly available Fable 5.1 and Mythos 5.1 — a variant with looser restrictions, available exclusively to verified organizations in the Project Glasswing program (cybersecurity and life sciences). Fable 5.1 doubles its predecessor's score in the agentic scientific research benchmark Terminal-Bench-Science (52.6% vs 24.7%), reduces cache read cost by 75%, and effectively lowers agentic workflow costs by 25–45%. This signals that competition is shifting from benchmark scores alone to operational economics and distribution models.
OpenAI confirms: Astra model requires stronger safeguards — it's too capable to release immediately
OpenAI revealed that the upcoming Astra model achieves significantly higher rates of arbitrary code execution than the public GPT-5.6 Sol, including the use of two zero-days in exploit chains. Sam Altman announced the model will launch "soon," but only after additional security layers are implemented. This is the first such explicit case where a manufacturer declares that a model has critical cyber capabilities and requires controlled deployment — strengthening the debate around external audits and regulations.
World Labs (Fei-Fei Li) shows Atlas — an omnimodel that combines 3D video and scene generation
World Labs has unveiled Atlas, a new omni world model pre-trained on text, images, video, and 3D data. It is a multimodal autoregressive diffusion transformer that generates video with pixel-perfect camera control (up to 1440p, 60 seconds) while simultaneously reconstructing scenes in 3D as point clouds and Gaussian splats. The model enters early access and merges two previously separate domains — generative video and 3D reconstruction — into a single step toward spatial intelligence.
Brussels sends first formal requests under the AI Act — clash with the US at G20
The European Commission confirmed it has sent the first formal requests for information to over 30 companies building general-purpose models (OpenAI, Google, Anthropic) — a step before potential proceedings for violating the AI Act. The questions concern model safety, independent audits, and post-deployment monitoring. On the same day at the G20 in Chapel Hill, Americans pushed for light-touch regulation (Carolina Principles), while Elon Musk criticized EU policy as hindering progress.
Higgsfield launches Genjutsu — video-to-video without prompts, with motion transfer and object swapping
Higgsfield has released Genjutsu, a video-to-video model with two features: Motion Transfer transfers motion from any clip to a new scene, and Object Swap replaces a specified element in the frame while preserving the rest of the shot. Everything works without text prompts — just a recording up to 30 seconds and up to 30 reference photos. Prices start at around $2 USD for a 15-second clip in 480p, eliminating costly reshoots in advertising and film production.
From the tech world
Critical vulnerability in Ruflo — CVSS 10.0 in AI agent management platform
Researchers at Noma Labs discovered a vulnerability (CVSS 10.0) in the open-source platform Ruflo (67k+ stars on GitHub), in the MCP Bridge component — an Express.js server exposing 233 tools, including shell and database access. The flaw allowed remote code execution and complete takeover of the AI agent environment.
TerminalFix — fake Cloudflare CAPTCHAs spreading a backdoor with a reverse tunnel
The TerminalFix malware campaign uses fake CAPTCHA pages impersonating Cloudflare to trick victims into clicking (ClickFix), triggering DLL sideloading and a Python backdoor with a reverse tunnel. The backdoor allows tunneling of arbitrary TCP traffic, bypassing firewalls and providing access to the victim's internal network.
Tip of the day
How to test AI models for safety before deploying them to an agent
Seeing the stories of Mythos 5.1 and Astra — models with advanced agentic capabilities can execute code and access your system before you realize something went wrong. Instead of relying solely on manufacturer promises, it's worth implementing a simple testing protocol: run the model in an isolated Docker container with no network access and give it a set of tasks requiring file access, configuration changes, or communication with external APIs. Check whether the model attempts to go beyond permitted boundaries. A concrete step: before you integrate an agent based on Fable 5.1 or any model into a production workflow, create a set of test prompts that check whether the model tries to run system commands, change environment variables, or connect to an unknown host. If it does — at least you know you need an additional policy layer (such as Google SA Key constraints or a Vault Agent sidecar) before granting it real permissions.
Reading list
Anthropic launches Fable 5.1 and Mythos 5.1 — a twin launch in the shadow of security
A full analysis of the launch from The Verge — why Anthropic decided on two versions of the same model and what it means for the future of AI distribution.
Critical Ruflo vulnerability — what every AI agent developer should know
Sekurak breaks down the flaw in MCP Bridge — a must-read if you work with agent orchestration platforms.
The AI world is no longer divided into better and worse models — but into those we trust, and those we must control.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.