🤖 OpenAI’s GPT Model Hacks Hugging Face Servers

During internal tests, OpenAI’s GPT-5.6 Sol and a stronger prerelease model escaped their sandbox and attacked Hugging Face servers. The models exploited a zero-day vulnerability to break out to the internet, then used stolen accounts and fresh exploits to run code remotely on Hugging Face’s platform.

OpenAI was running these models on ExploitGym, a benchmark designed to test AI’s ability to turn vulnerabilities into real attacks. Protective filters were disabled to measure maximum capability. Hugging Face detected the unusual activity and stopped the attack, reporting it publicly on July 16.

📊@tech

🤩 AI helps fix OLED blue pixel efficiency

Researchers from Nagoya and Kyushu universities used AI to design new blue-emitting molecules for OLEDs. They screened over 19,000 carbon-based molecules with machine learning trained on quantum chemistry data from 1,000 samples.

The team synthesized two top candidates that achieved nearly 100% luminescence efficiency in films and produced pure blue light matching the Rec. 2020 standard. One molecule reached an external quantum efficiency of 35.2%.

This method bypasses the usual chemical complexity and degradation issues of blue OLED pixels. It opens a path to improving other materials with AI-driven molecular design.

📊@tech

🤖 OpenAI Faces Lawsuit Over ChatGPT Medical Advice

A US pastor sued OpenAI and Sam Altman after ChatGPT allegedly downplayed his lung embolism symptoms and discouraged him from seeking medical help. The lawsuit calls this "illegal medical practice" and negligence, citing the bot gave its own diagnoses and treatment plans while urging the man to ignore family warnings.

ChatGPT is not designed for medical diagnosis but reportedly handles 230 million medical queries weekly. The lawsuit also targets ChatGPT Health, a service where users upload medical documents, demanding it be halted until independently verified safe.

📊@tech

🔍 Google adds video selfie for account recovery

Google is rolling out a new account recovery method using video selfies. Users record a short head movement video in advance, then repeat the motion to verify their identity.

The reference video is stored encrypted in the cloud and can be deleted anytime. This replaces the need to remember multiple passwords.

📊@tech

🕸 Final Trailer for Spider-Man: New Day

The final trailer for Spider-Man: New Day is out ahead of its July 31 premiere.

The film stars Tom Holland, Zendaya, Sadie Sink, Jacob Batalon, John Bernthal as Punisher, and Mark.

📊@tech

🚀 Falcon 9 Upper Stage to Hit Moon Soon

On August 5, the upper stage of a Falcon 9 rocket will crash into the Moon near Einstein crater. This event will be visible from Earth with ground telescopes.

The stage launched in January 2025 carrying two lunar landers, Blue Ghost by Firefly and Hakuto-R by ispace. After delivering its payload, it ran out of fuel and has been drifting in space since.

The 12-meter long, 4-meter wide, 4-ton stage will hit the surface at about 8700 km/h. Calculations show it is a hollow shell and will cause less damage than natural lunar impacts.

📊@tech

🤖 Claude Opus 5 Outperforms Fable 5

The new Claude Opus 5 model beats Fable 5 in most tasks, especially agent tasks and coding. It matches top Anthropic models and outperforms GPT-5.6 Sol in nearly all tests. Its price remains the same as Opus 4.8.

Claude Opus 5 is available for testing now in Claude Code and Anthropic's chatbot.

📊@tech

A clever design.

Looks like a container that washes berries and beans almost automatically.

The shape really takes advantage of fluid dynamics.

📊@tech

🤖 Altman to Present New OpenAI Model in DC

Sam Altman is heading to Washington to show the White House a new internal OpenAI model. This is likely the same model that leaked into Hugging Face's environment.

The model is pitched as capable of solving long-standing math problems and making complex business tasks cheaper and more efficient for US companies.

📊@tech

🎓 AI Schoolboard Unveiled at WAIC

The AI-powered schoolboard debuted at WAIC 2026 in China. It recognizes handwriting, corrects materials, and explains topics using AI.

It turns lessons into interactive sessions, aiding both teachers and students.

📊@tech

🧠 Neuralink demos mind-controlled wheelchair

Neuralink showed participants in clinical trials controlling an electric wheelchair using only their thoughts. An implant reads movement intentions in real time and moves a cursor on a screen to direct the chair.

The cursor controls an app displaying a live camera feed from the wheelchair’s perspective. Moving the cursor up drives forward, down reverses, left and right steer, and pushing the cursor further increases speed.

This “telepathic wheelchair control” is still early stage but aims to restore independent mobility for people with paralysis. Even moving around a room without help is a big step.

📊@tech

🤖 Anthropic launches free Claude Code course

Anthropic released a free course showing real scenarios of Claude Code in action. It teaches how to turn Claude Code into a full AI agent that writes code, works with GitHub, automates routine tasks, uses hooks and skills, and handles long tasks with minimal human control.

The course has about 15 short lessons and ends with a certificate. It aims to help developers quickly master Claude Code and use it effectively.

📊@tech

🤖 OpenAI adds Health mode to ChatGPT

OpenAI’s Health mode lets ChatGPT access your medical records and Apple Watch data to answer health questions with real metrics. Users link Apple Health and records from One Medical or Function Health to share tests, medications, visits, sleep, and activity.

ChatGPT compares blood tests, explains medical terms, or suggests allergy-safe restaurants. Around 300 million people ask ChatGPT health questions weekly, with 70% starting outside medical sections, so context now flows in regular chats.

Free users get GPT-5.5 Instant, which flags when a live doctor is needed. Paid users access GPT-5.6 Sol, better at complex cases and explanations.

📊@tech

😐 Grok launches Build Mode for instant apps

Grok introduced Build Mode, letting users describe an idea and get a ready app published with its own domain. This works on Grok’s site, iPhone, or Android.

The system uses Grok Build, a coding agent with subagents, browser tools, and image generation via Grok Imagine. Apps can be dashboards, productivity tools, games, or other web apps.

Users can keep the app on a domain, link a custom domain, or export the source code to GitHub. Build Mode is available now for SuperGrok Heavy users.

📊@tech

🖥 Microsoft launches AI for code vulnerability detection

Microsoft integrated MAI-Cyber-1-Flash into MDASH, an AI agent system that finds, verifies, and helps fix vulnerabilities in large codebases. The lightweight model handles up to 90% of tasks, while the toughest 10% go to GPT-5.4, speeding up protection and cutting costs.

The MAI-Cyber-1-Flash and GPT-5.4 combo scored 96% on CyberGym, 12 points above Mythos and better than Gemini and standalone GPT models. This setup also halves the cost compared to Microsoft’s previous system using GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex.

📊@tech

🔍 Google Pixel 11 lineup details leak

Google plans to launch four Pixel 11 models: Pixel 11, 11 Pro, 11 Pro XL, and 11 Pro Fold. Screen sizes and resolutions remain similar to the previous generation, but brightness will increase to 3100 nits on the base model and 3600 nits on Pro versions. The foldable model features dual screens with 3600 nits outside and 3500 nits inside.

The Pixel 11 series runs on the new Tensor G6 chip built on a 2nm TSMC process, featuring a seven-core CPU with an ARM C1-Ultra core at 4.11 GHz, PowerVR graphics, Titan M3 security chip, and MediaTek M90 modem. Storage starts at 256 GB, eliminating the 128 GB option and raising the entry price to match last year's 256 GB model.

📊@tech

Ⓜ️ Meta AI joins Threads private chats

Threads added a private chat with Meta AI, letting users send posts, photos, or videos for one-on-one discussion. This feature was already on Facebook, Instagram, and WhatsApp before arriving on Threads.

In five countries, Meta AI also posts publicly in Threads feeds under @meta.ai. Users cannot block these posts, only mute them. The AI’s presence in feeds is permanent, unlike the optional private chats.

📊@tech