😆 Just looking around

Researchers say OpenAI agents accessed two Hugging Face accounts and probed the platform for weaknesses nearly two months before the major July breach.

↖️ https://durovscode.com/openai-agents-hugging-face-may-probing

💅 No scrolling until 13

The European Commission wants to ban children under 13 from social media across the EU. The plan also targets addictive features, online games and AI chatbots.

More on the proposal:

↖️ https://durovscode.com/eu-social-media-ban-under-13

😄 The models had other plans

OpenAI says the industry still hasn’t solved how to reliably keep advanced AI models in check. It has launched a public reporting system and released its first six reports.

During training and evaluation, models hid mistakes from users, used someone else’s exposed API key without permission, and uploaded task files to public websites.

The six cases and OpenAI’s new reporting rules:

↖️ https://durovscode.com/openai-details-six-cases-of-ai-models-hiding-errors-and-taking-unauthorized-actions

🤐 A little help from the competition

Researchers used Anthropic’s Claude to breach OpenAI employee accounts and demonstrate access to its private code repository.

Full story:

↖️ https://durovscode.com/claude-openai-hacktron-hack

🤩 Bad optics

Australia is considering banning camera-equipped smart glasses from federal offices and service centers.

What would change:

↖️ https://durovscode.com/australia-smart-glasses-ban

🤔 Teamwork, apparently

The US has proposed a “notification mechanism” with China for AI incidents that could affect national security.

More on the proposal:

↖️ https://durovscode.com/us-china-ai-incident-alerts

😄 Financial advice, apparently

Saturn tested 18 popular AI models on personal finance questions. On average, ChatGPT, Claude, Gemini, Copilot and Grok gave incorrect answers 57% of the time. On more complex tasks, the error rate jumped to 88%.

In one case, incorrect tax advice from Claude could potentially have cost a user an extra £17,500.

Which models performed best — and worst:

↖️ https://durovscode.com/chatgpt-claude-gemini-personal-finance-errors

😏 Size matters, apparently

Apple’s anniversary iPhone 20 Pro models may get the first increase in Pro screen sizes since 2024. The company is also reportedly testing a nearly borderless design curved on all four sides.

More on the leak:

↖️ https://durovscode.com/iphone-20-pro-display-sizes-borderless-design

😆 A costly detour

Ireland fined Google €403 million over its handling of users’ location data. The regulator found that Google kept some of it longer than necessary and failed to clearly explain how people’s whereabouts could be used.

What the ruling says:

↖️ https://durovscode.com/google-403-million-location-data-fine

🦴 Grok 4.7 goes to work

SpaceXAI has released Grok 4.7, its new flagship model built for long-running agentic tasks. On some benchmarks, it also outperforms GPT-6 Astra.

API pricing stays unchanged from Grok 4.6, starting at $2 per million tokens.

Benchmarks, capabilities and availability:

↖️ https://durovscode.com/spacexai-grok-4-7-coding-ai-model

🛸 Permission slip

A researcher found a flaw in Meta’s Muse for Mac that could let malware already on your computer take control of the AI assistant and use its access to your files and connected accounts.

How the attack works:

↖️ https://durovscode.com/meta-muse-mac-security-flaw

😄 Because video messages needed 60 FPS

Telegram is testing more advanced video message controls on Android, with options to:

📌 record at 30 or 60 FPS
📌 choose between 360p and 480p
📌 adjust camera quality and bitrate

More on what Telegram is testing:

↖️ https://durovscode.com/telegram-video-messages-60-fps-android