2 min read
In short
Today’s roundup covers OpenAI’s new content requirements and risks for agentic systems. We’ll also take a look at models for local AI and practical materials for testing agents.
Today’s roundup covers new content requirements from OpenAI and risks for agentic systems. We’ll also look at models for local AI and practical materials on testing agents.
🔥 Hot:
🔹 OpenAI will start labeling ChatGPT and Codex texts in the EU: This is how the company intends to comply with the AI Act, although editing the text may make the labels harder to detect. 🔹 MCP vulnerabilities allow one AI agent to attack another: This issue could lead to database leaks and the exposure of other sensitive information. 🔹 Reflection introduced the open Beam model, targeting lower computing costs: The company is offering organizations the opportunity to create local AI systems and fine-tune them on their own data.
➡️ News:
🔹 Nolla Health launched a pilot for AI acne prescriptions in Utah: The app evaluates breakouts from facial photos and issues prescriptions on its own, while physician oversight is gradually being reduced. 🔹 Google may expand Gemini Call for Me to calls with loved ones: A discovered screen for the feature includes examples of calling a mom and friends to pass along a message. 🔹 A command-line utility removes Apple Intelligence from macOS 27: It helps disable AI features and free up to 12 GB, since macOS 27 has no switch for this.
➡️ Useful materials:
🔹 Analysis: how AI labs claim breakthroughs in mathematics: The article discusses high-profile solutions to complex problems and how the race between labs is affecting the discipline itself. 🔹 How to test agents that call functions: The article examines an error in which an agent acts without the necessary condition, reports success, and slips past conventional checks. 🔹 Why voice agents make mistakes when reading dates and numbers: Even good speech synthesis cannot help when an agent pronounces numbers incorrectly during a phone conversation. 🔹 How ELMA365 automated testing in three iterations: The team explains how they introduced AI to check forms, roles, approvals, and scenarios.
📝 If you’d like to add other news and materials to the list, write in the comments.