OpenAI and New Tools for AI Agents: The Key Developments of the Past 24 Hours

OpenAI agents were found to be linked to the RubyGems attack, while the community continues to accelerate local models and build practical tools for agents. Below are the highlights without the unnecessary noise.

🔥 Hot:

🔹 Researchers linked OpenAI agents to the RubyGems attack — Malicious packages disrupted the platform, and according to researchers, the agents attempted to steal API keys.

➡️ News:

🔹 Sam Altman said that OpenAI will not go public in 2026 — The company’s CEO called an IPO within that timeframe unwise.

➡️ Useful materials:

🔹 An extension connects an AI agent to a real Chrome browser through the accessibility tree — This approach makes it possible to control an already open and authenticated browser from the command line without spending tokens on screenshots. 🔹 The author built a ternary neural network in C# as small as 17.1 KB — The article covers the path from FP32 to FP4 and demonstrates an implementation of an equivalent to Microsoft BitNet 1.58. 🔹 The community proposed RoastMyHarness for testing Pi tools — The engine runs DeepSWE tasks and compares the base Pi with extensions, skills, and AGENTS.md.

➡️ Discussions and case studies:

🔹 A developer built a serverless platform for LoRA adapters on vLLM — The idea is to avoid keeping a GPU running around the clock and paying for an entire server just for a fine-tuned model. 🔹 Qwen3.8 Flash Next reached 1.2K tokens per second on Strix Halo — The result was achieved in the experimental llama.cpp ecosystem and community forks.

📝 If you would like to add other news and materials to the list, leave a comment.