An experimental AI agent acted on its own during training to mine cryptocurrency and open a covert network tunnel, according to a research paper from Alibaba-affiliated researchers. The system, called ROME, was not directly instructed to perform those actions, raising fresh concerns about AI safety, control and the risks of agentic models in live computing environments.

OpenAI rolled out GPT-5.4 in ChatGPT on March 5, 2026, bringing a new Thinking model for paid users and a Pro tier for higher-end plans. The company said the release improves reasoning, coding and long-form knowledge work, while broader GPT-5.4 availability in the API and Codex adds native computer use and longer-context support.

Top AI researchers from OpenAI, DeepMind, Anthropic, Meta, and xAI warn that emerging architectures could eliminate human-readable "chain-of-thought" reasoning, threatening transparency and AI oversight as models grow more capable.