❌

Vue lecture

Il y a de nouveaux articles disponibles, cliquez pour rafraîchir la page.

OBS Studio 32 débarque avec un tout nouveau moteur de rendu pour macOS

✇Korben
Korben

Passer d'OpenGL à Metal, c'était visiblement pas une mince affaire pour l'équipe d'OBS. La techno d'Apple est sortie y'a 10 ans sous macOS mais ça leur a pris un peu de temps pour la migration... Et n'allez pas croire que je "juge"... Tout ce temps, c'est normal car c'est un soft multiplateforme, donc faut gérer trois écosystèmes en parallèle et ça, ça prend un temps fou.

Tous les effets visuels d'OBS ont dû être réécrits pour fonctionner avec Metal, le langage graphique d'Apple étant bien plus exigeant que celui de Windows et la preview peut parfois légèrement saccader à cause de macOS, mais le flux final reste impeccable.

Niveau performances, Metal fait aussi bien voire mieux qu'OpenGL dans les builds Release mais c'est surtout pour le débogage que ça change tout car les développeurs ont maintenant accès à des outils de diagnostic bien plus performants, ce qui devrait accélérer les corrections de bugs et les futures améliorations.

Pour l'activer (ouais on est chaud !!), c'est hyper simple. Vous allez dans les paramètres d'OBS 32.0, onglet Avancé, section Vidéo, et vous sélectionnez Metal dans le menu déroulant du renderer. Un petit redémarrage de l'appli et hop, vous êtes passé sur le nouveau moteur.

Ce qui est cool aussi avec cette version 32.0, c'est qu'elle inclut un gestionnaire de plugins et des améliorations pour les fonctionnalités NVIDIA RTX.

L'équipe OBS bosse aussi sur des backends Vulkan pour Linux et Direct3D 12 pour Windows, parce que les anciennes APIs comme OpenGL et D3D11 reçoivent de moins en moins de support des fabricants de GPU, donc si vous êtes sur Linux ou Windows, votre tour viendra aussi.

Voilà, après si ça bug, revenez sur OpenGL, mais y'a quand même de bonnes chances que ça tourne mieux qu'avant.

Source

Quand une caméra de surveillance TP-Link laisse traîner ses clés HTTPS partout...

✇Korben
Korben

Vous avez peut-être une caméra Tapo C200 qui tourne chez vous pour surveiller le chat, le bébé ou l'entrée. C'est mon cas et j'adore cette caméra mais j'ai une mauvaise nouvelle à vous annoncer... Le chercheur en sécurité Simone Margaritelli (alias evilsocket) vient de passer 150 jours à la disséquer et le résultat n'est pas glorieux pour TP-Link.

Alors déjà, commençons par le plus gros WTF qu'il a découvert... la clé privée HTTPS de la caméra, ce truc censé être ultra-secret qui permet de chiffrer les communications. Et bien elle est hardcodée dans le firmware. C'est donc la même clé pour TOUTES les caméras du même modèle. Du coup, n'importe qui peut faire un Man-in-the-Middle et intercepter ce que vous voyez sur votre caméra. Ah on se met bien déjà là, hein ? ^^

Et attendez, ça ne s'arrête pas là puisque Margaritelli a trouvé un bucket S3 chez Amazon, totalement ouvert au public, qui contient TOUS les firmwares de TOUS les produits TP-Link. C'est open bar, sans authentification, Noël avant l'heure pour les chercheurs en sécu... et les hackers.

En fouillant le firmware avec Ghidra et Claude (oui, l'IA a aidé au reverse engineering), le chercheur a découvert quatre failles critiques. La première, c'est un buffer overflow dans le parser SOAP XML utilisé par le protocole ONVIF. En gros, si vous envoyez un message trop long, la caméra plante. Pas besoin d'être authentifié pour ça, une requête HTTP suffit.

La deuxième faille est du même genre mais dans le header Content-Length. Envoyez 4294967295 (le max d'un entier 32 bits) et boum, integer overflow. Et la troisième, c'est la cerise sur le gâteau puisque l'endpoint connectAp reste accessible sans authentification même après le setup initial. Du coup, un attaquant peut forcer votre caméra à se connecter à son propre réseau WiFi malveillant et intercepter tout le flux vidéo. Vous ne vous y attendiez pas à celle-là, si ?

Et la quatrième faille, oubliée nulle part ailleurs c'est l'API scanApList qui balance la liste de tous les réseaux WiFi autour de la caméra, sans auth. Avec les BSSID récupérés et un outil comme apple_bssid_locator, on peut géolocaliser physiquement la caméra à quelques mètres près. Sur les 25 000 caméras exposées sur le net, ça fait froid dans le dos.

Le plus frustrant dans cette histoire, c'est que Margaritelli a signalé tout ça en juillet 2025 et TP-Link a demandé des rallonges de délai, encore et encore, durant plus de 150 jours. Et au final, les failles ont été corrigées mais pas de patch sur les pages publiques des CVE. Ah et petit détail rigolo, comme TP-Link est sa propre autorité de numérotation CVE, ils s'auto-évaluent sur leurs propres failles. Donc y'a pas de conflit d'intérêt du tout... ahem ahem...

Le chercheur estime qu'environ 25 000 de ces caméras sont exposées directement sur Internet donc si comme moi, vous en avez une, vérifiez que le firmware est bien à jour et surtout, ne l'exposez JAMAIS directement sur le net. Mettez-la derrière un VPN ou un réseau isolé.

Je trouve ça cool que Margaritelli ait utilisé de l'IA pour accélérer la phase de reverse engineering. Avec Claude Opus et Sonnet avec GhidraMCP, il a pu analyser le code assembleur et c'est comme ça que l'IA a identifié rapidement les fonctions vulnérables et expliqué le fonctionnement du code. Bref, l'IA comme outil de hacking, c'est assez ouf...

Voilà, donc si vous avez du matos TP-Link chez vous, gardez un œil sur les mises à jour et réfléchissez à deux fois avant de l'exposer sur le net. Et si vous aimez la lecture, l'analyse complète est dispo sur le blog d'evilsocket .

Beau boulot !

This $2,899 Desktop AI Computer With RTX 5090M Lets You Cancel Every AI Subscription Forever

✇Yanko Design
Sarang Sheth

Look across the history of consumer tech and a pattern appears. Ownership gives way to services, and services become subscriptions. We went from stacks of DVDs to streaming movies online, from external drives for storing data and backups to cloud drives, from MP3s on a player to Spotify subscriptions, from one time software licenses to recurring plans. But when AI arrived, it skipped the ownership phase entirely. Intelligence came as a service, priced per month or per million tokens. No ownership, no privacy. Just a $20 a month fee.

A device like Olares One rearranges that relationship. It compresses a full AI stack into a desktop sized box that behaves less like a website and more like a personal studio. You install models the way you once installed apps. You shape its behavior over time, training it on your documents, your archives, your creative habits. The result is an assistant that feels less rented and more grown, with privacy, latency, and long term cost all tilting back toward the owner.

Designer: Olares

Click Here to Buy Now: $2,899 $3,999 (28% off) Hurry! Only 15/320 units left!

The pitch is straightforward. Take the guts of a $4,000 gaming laptop, strip out the screen and keyboard, put everything in a minimalist chassis that looks like Apple designed a chonky Mac mini, and tune it for sustained performance instead of portability. Dimensions are 320 x 197 x 55mm, weighs 2.15 kg without the PSU, and the whole package pulls 330 watts under full load. Inside sits an Intel Core Ultra 9 275HX with 24 cores running up to 5.4 GHz and 36 MB of cache, the same chip you would find in flagship creator laptops this year. The GPU is an NVIDIA GeForce RTX 5090 Mobile with 24 GB of GDDR7 VRAM, 1824 AI TOPS of tensor performance, and a 175W max TGP. Pair that with 96 GB of DDR5 RAM at 5600 MHz and a PCIe 4.0 NVMe SSD, and you have workstation level compute in a box smaller than most soundbars.

Olares OS runs on top of all that hardware, and it is open source, which means you can audit the code, fork it, or wipe it entirely if you want. Out of the box it behaves like a personal cloud with an app store containing over 200 applications ready to deploy with one click. Think Docker and Kubernetes, but without needing to touch a terminal unless you want to. The interface looks clean, almost suspiciously clean, like someone finally asked what would happen if you gave a NAS the polish of an iPhone. You get a unified account system so all your apps share a single login, configurable multi factor authentication, enterprise grade sandboxing for third party apps, and Tailscale integration that lets you access your Olares box securely from anywhere in the world. Your data stays on your hardware, full stop.

I have been tinkering with local LLMs for the past year, and the setup has always been the worst part. You spend hours wrestling with CUDA drivers, Python environments, and obscure GitHub repos just to get a model running, and then you realize you need a different frontend for image generation and another tool for managing multiple models and suddenly you have seven terminal windows open and nothing talks to each other. Olares solves that friction by bundling everything into a coherent ecosystem. Chat agents like Open WebUI and Lobe Chat, general agents like Suna and OWL, AI search with Perplexica and SearXNG, coding assistants like Void, design agents like Denpot, deep research tools like DeerFlow, task automation with n8n and Dify. Local LLMs include Ollama, vLLM, and SGIL. You also get observability tools like Grafana, Prometheus, and Langfuse so you can actually monitor what your models are doing. The philosophy is simple. String together workflows that feel as fluid as using a cloud service, except everything runs on metal you control.

Gaming on this thing is a legitimate use case, which feels almost incidental given the AI focus but makes total sense once you look at the hardware. That RTX 5090 Mobile with 24 GB of VRAM and 175 watts of power can handle AAA titles at high settings, and because the machine is designed as a desktop box, you can hook it up to any monitor or TV you want. Olares positions this as a way to turn your Steam library into a personal cloud gaming service. You install your games on the Olares One, then stream them to your phone, tablet, or laptop from anywhere. It is like running your own GeForce Now or Xbox Cloud Gaming, except you own the server and there are no monthly fees eating into your budget. The 2 TB of NVMe storage gives you room for a decent library, and if you need more, the system uses standard M.2 drives, so upgrades are straightforward.

Cooling is borrowed from high end laptops, with a 2.8mm vapor chamber and a 176 layer copper fin array handling heat dissipation across a massive 310,000 square millimeter surface. Two custom 54 blade fans keep everything moving, and the acoustic tuning is genuinely impressive. At idle, the system sits at 19 dB, which is whisper quiet. Under full GPU and CPU load, it climbs to 38.8 dB, quieter than most gaming desktops and even some laptops. Thermal control keeps things stable at 43.8 degrees Celsius under sustained loads, which means you can run inference on a 70B model or render a Blender scene without the fans turning into jet engines. I have used plenty of small form factor PCs that sound like they are preparing for liftoff the moment you ask them to do anything demanding, so this is a welcome change.

RAGFlow and AnythingLLM handle retrieval augmented generation, which lets you feed your own documents, notes, and files into your AI models so they can answer questions about your specific data. Wise and Files manage your media and documents, all searchable and indexed locally. There is a digital secret garden feature that keeps an AI powered local first reader for articles and research, with third party integration so you can pull in content from RSS feeds or save articles for later. The configuration hub lets you manage storage, backups, network settings, and app deployments without touching config files, and there is a full Kubernetes console if you want to go deep. The no CLI Kubernetes interface is a big deal for people who want the power of container orchestration but do not want to memorize kubectl commands. You get centralized control, performance monitoring at a glance, and the ability to spin up or tear down services in seconds.

Olares makes a blunt economic argument. If you are using Midjourney, Runway, ChatGPT Pro, and Manus for creative work, you are probably spending around $6,456 per year per user. For a five person team, that balloons to $32,280 annually. Olares One costs $2,899 for the hardware (early-bird pricing), which breaks down to about $22.20 per month per user over three years if you split it across a five person team. Your data stays private, stored locally on your own hardware instead of floating through someone else’s data center. You get a unified hub of over 200 apps with one click installs, so there are no fragmented tools or inconsistent experiences. Performance is fast and reliable, even when you are offline, because everything runs on device. You own the infrastructure, which means unconditional and sovereign control over your tools and data. The rented AI stack leaves you as a tenant with conditional and revocable access.

Ports include Thunderbolt 5, RJ45 Ethernet at 2.5 Gbps, USB A, and HDMI 2.1, plus Wi-Fi 7 and Bluetooth 5.4 for wireless connectivity. The industrial design leans heavily into the golden ratio aesthetic, with smooth curves and a matte aluminum finish that would not look out of place next to a high end monitor or a piece of studio equipment. It feels like someone took the guts of a $4,000 gaming laptop, stripped out the compromises of portability, and optimized everything for sustained performance and quietness. The result is a machine that can handle creative work, AI experimentation, gaming, and personal cloud duties without breaking a sweat or your eardrums.

Olares One is available now on Kickstarter, with units expected to ship early next year. The base configuration with the RTX 5090 Mobile, Intel Core Ultra 9 275HX, 96 GB RAM, and 2 TB SSD is priced at a discounted $2,899 for early-bird backers (MSRP $3,999). That still is a substantial upfront cost, but when you compare it to the ongoing expense of cloud AI subscriptions and the privacy compromises that come with them, the math starts to make sense. You pay once, and the machine is yours. No throttling, no price hikes, no terms of service updates that quietly change what the company can do with your data. If you have been looking for a way to bring AI home without sacrificing capability or convenience, this is probably the most polished attempt at that idea so far.

Click Here to Buy Now: $2,899 $3,999 (28% off) Hurry! Only 15/320 units left!

The post This $2,899 Desktop AI Computer With RTX 5090M Lets You Cancel Every AI Subscription Forever first appeared on Yanko Design.

This $7,000 Robot Shapeshifts Into 3 Different Machines

✇Yanko Design
Ida Torres

Imagine a robot that can transform like a high-tech LEGO set, swapping out legs for arms or wheels depending on what the day throws at it. That’s exactly what LimX Dynamics has cooked up with their latest creation, the Tron 2, and honestly, it’s making me rethink everything I thought I knew about what robots can do.

The Tron 2 isn’t your typical one-trick-pony robot. This thing is basically the Swiss Army knife of the robotics world. Chinese startup LimX Dynamics just unveiled this modular marvel that can morph between three completely different configurations: a dual-armed humanoid torso, a wheeled-leg explorer, or a bipedal walker that can actually climb stairs without making you nervous. And get this, you can switch between these forms with just a screwdriver. No fancy tools, no complicated procedures. Just some strategic unscrewing and you’ve got a whole new robot.

Designer: LimX Dynamics

The company’s demo video starts with something delightfully surreal: just a pair of robotic legs casually strolling along, completely headless and armless. Then, like watching a transformer come to life in real time, those same leg components get repurposed into arms, complete with a head and torso. Suddenly, you’ve got a full humanoid lifting heavy water bottles and showing off its surprisingly impressive strength.

What makes the Tron 2 particularly fascinating is its intelligence layer. This isn’t just a mechanical chameleon. It’s powered by advanced AI and built on what’s called a vision-language-action platform, which essentially means it can see, understand commands, and actually do something useful with that information. The robot comes with a fully open software development kit that plays nice with both ROS1 and ROS2, making it a dream for researchers and developers who want to experiment without fighting proprietary systems.

Performance-wise, the specs are genuinely impressive. Each of its dual arms features seven degrees of freedom with a reach of 70 centimeters and can handle up to 10 kilograms of payload together. The wheeled configuration offers about four hours of runtime and can haul around 30 kilograms of cargo, while the bipedal mode excels at navigating tricky terrain like staircases that would leave most wheeled robots stuck at the bottom. The demo footage shows Tron 2 doing things that feel almost show-offy: playing table tennis, performing cartwheels, rolling around smoothly on wheels, and conquering staircases with the confidence of someone who’s done it a thousand times. It’s the kind of versatility that makes you wonder why we’ve been so committed to single-purpose robots for so long.

And here’s where things get really interesting. LimX is positioning the Tron 2 as ideal for future Mars missions. Think about it: on Mars, you can’t exactly call a repair truck when something breaks or send a specialized robot for every different task. You need something adaptable, something that can switch roles as mission needs evolve. The modular design means you could potentially swap out damaged components or reconfigure for different tasks without needing an entirely new robot shipped from Earth.

For research labs, the Tron 2 offers something that’s been surprisingly rare: a flexible test bed that can support multiple types of projects without requiring a whole fleet of different robots. Whether you’re studying manipulation, locomotion, or AI integration, you can configure the same platform to suit your specific needs. Perhaps most surprisingly, this technological marvel starts at just 49,800 Chinese yuan, which translates to around $7,000 USD. For context, that’s dramatically cheaper than many specialized robots that can only do a fraction of what the Tron 2 offers. Pre-orders are already open, though LimX hasn’t fully disclosed all the pricing details or specified exactly who their target customers are.

The Tron 2 represents something bigger than just another cool robot demo. It’s pointing toward a future where adaptability matters more than specialization, where one well-designed platform can handle whatever challenges come its way. Whether it ends up exploring Mars or revolutionizing warehouse operations here on Earth, this shape-shifting bot is definitely one to watch.

The post This $7,000 Robot Shapeshifts Into 3 Different Machines first appeared on Yanko Design.

❌