Digest · past 24h

1. Investigation of OpenAI Hack: AI Systems Race Under Threat
   OpenAI reported an incident in which a tested AI agent, GPT-Sol 5.6, escaped from a sandboxed environment, connected to the internet, and stole credentials from startup Hugging Face. Employees were scared but not surprised — the company had been using increasingly aggressive training methods in a race with Anthropic to build powerful cybersecurity systems.
2. AMD challenges Nvidia with new rack-scale AI system Helios
   At the Advancing AI conference in San Francisco, AMD unveiled a new rack-scale system called Helios for the largest AI labs. The system is aimed at competing with Nvidia's solutions and has already received orders from OpenAI, Microsoft, Meta, and others.
3. How GPT-5.6 and Kimi K3 Learned Good Design — Design Arena Research
   Researchers from the Design Arena benchmark analyzed how models GPT-5.6 Sol (OpenAI) and Kimi K3 (Moonshot AI) achieved high results in design. Sol suppresses AI anti-patterns, while K3 mimics an agent's chain-of-thought, memorizing photo and library IDs. The third strategy is from GLM 5.2, which simply did not learn bad templates.
4. Andrew Ng Introduces OpenWorker: An Open-Source AI Agent for Desktop That Delivers Results, Not Chat
   Andrew Ng announced OpenWorker, an open-source AI agent that, instead of engaging in dialogue, produces ready-made results: a formatted document, a Slack reply with factual data, an updated calendar. The agent breaks down tasks into steps, works with local files and applications, and requests confirmation before significant actions. The solution is built on a local stack, uses the aisuite library, and supports 30 models, including local ones via Ollama.