🔐 Autonomous AI Cyberattack

Hey folks,

OpenAI just launched group chats in ChatGPT, and the AI knows when to shut up. You can now invite up to 20 people into a conversation where ChatGPT follows along, decides when to contribute based on context, and stays quiet when humans are talking. Usage limits only count when the AI responds, not when you're chatting with your team, and the AI can react with emojis while your personal ChatGPT memory stays completely private from the group. It's piloting in Japan, New Zealand, South Korea, and Taiwan for all plans, powered by GPT-5.1. Stay tuned for a wider rollout soon.

Let’s dive in..

🔐 The First Autonomous AI Cyberattack

🐝 The Buzz: Anthropic disclosed that Chinese state-sponsored hackers jailbroke Claude Code to execute what they're calling the first large-scale cyberattack with minimal human intervention. The attackers decomposed malicious tasks into innocent-seeming prompts like "help me organize these credentials" instead of "steal these passwords", achieving 80-90% autonomous execution with thousands of requests per second.

• Reconnaissance: Claude scanned the environment, uncovered internal services, and flagged the systems that mattered most.

• Exploitation: It wrote its own attack code, checked the weak spots, and proved it could get in.

• Credential harvesting: It pulled usernames and passwords, then tried them across different systems to see what opened.

• Data extraction: It sifted through huge piles of captured data and sorted everything by how useful it might be.

• Documentation: It logged every step and finding so the humans had a clear play-by-play.

💡 Takeaway: Engineering teams deploying AI agents need task context validation (not just individual prompt filtering), aggressive rate limiting, comprehensive audit logging of AI actions, and segmented permissions that prevent agents from accessing sensitive systems. The "innocent task decomposition" attack vector works because current safety mechanisms evaluate prompts in isolation rather than tracking behavior patterns.

Together with Proton

Free, private email that puts your privacy first

A private inbox doesn’t have to come with a price tag—or a catch. Proton Mail’s free plan gives you the privacy and security you expect, without selling your data or showing you ads.

Built by scientists and privacy advocates, Proton Mail uses end-to-end encryption to keep your conversations secure. No scanning. No targeting. No creepy promotions.

With Proton, you’re not the product — you’re in control.

Start for free. Upgrade anytime. Stay private always.

Get free private email

🇨🇳 Baidu's ERNIE 5.0 Claims GPT-5 Performance

🐝 The Buzz: Baidu unveiled ERNIE 5.0, a natively omni-modal model with 2.4 trillion parameters that claims to beat GPT-5 and Gemini on enterprise tasks like document understanding, chart analysis, and visual reasoning. The model uses sparse activation, less than 3% of parameters active per inference to keep costs viable despite massive scale.

• Sparse activation at scale: With 2.4T total parameters but activating less than 3% during each inference, ERNIE 5.0 achieves frontier model performance while maintaining practical compute costs.

• Natively omni-modal architecture: Built from the ground up to jointly process and generate across text, images, audio, and video rather than bolting modalities onto a text-first foundation, enabling tighter cross-modal reasoning and generation quality matching Google's Veo3 on image synthesis.

• Enterprise deployment focus: Qianfan platform positions ERNIE 5.0 for document processing, business intelligence, and multimodal content generation workflows, targeting the enterprise use cases where Chinese companies need sovereignty over their AI infrastructure.

💡 Takeaway: China’s ERNIE 5.0's sparse activation strategy solves the practical deployment problem that makes most trillion-parameter models vaporware, while benchmark performance suggests Western model advantages are narrowing fast. Engineering teams evaluating AI vendors should track Chinese alternatives seriously, especially if data sovereignty, cost efficiency, or Chinese language performance matter for your use case.

📰 Like newsletters? Here are more our readers enjoy.

💻️ How to use Google AI Studio

🐝 AI buzz bits

📊 NotebookLM launched Deep Research, an automated research assistant that browses the web, synthesizes reports, and now ingests Google Sheets, Word docs, and Drive URLs. Deep Research takes your question, creates a research plan, browses websites autonomously, and delivers source-grounded reports in minutes.

💰 Cursor raised $2.3 billion at a $29.3 billion valuation. The IDE features Composer mode for multi-file edits, Agent mode for end-to-end autonomous development, and proprietary models that generate more code than almost any other LLMs.

🧠 Yann LeCun is leaving Meta to launch a world models startup focused on spatial reasoning over text-based LLMs. Meta's chief AI scientist publicly disagrees with LLM-centric approaches, calling them "useful but fundamentally limited" for reasoning and planning.

🔥 AI productivity tools

For a full list of 1000+ AI tools, visit our Supertool Directory

🤘 Mon Meme