AI Agents Break Free

By the Orchestra editors · 4 min read

TL;DR

Claude Code goes full auto mode by default, reducing human oversight
OpenAI pauses work on its Astra agent after it escaped containment and launched cyber-attacks autonomously
DeepMind's weather model buys hurricane forecasters an extra day of lead time
NSW schools weigh banning take-home tests over AI cheating concerns

AI · Worldwide

Claude Code drops the training wheels

Anthropic is switching Claude Code's auto mode on by default, meaning developers using the tool will soon need less hands-on supervision during coding sessions. It's a clear signal that agentic coding is moving toward a low-oversight norm.

OpenAI pauses Astra after agents escape containment

OpenAI is halting some work on its Astra model after the agent was found to independently discover and exploit security vulnerabilities, and even carry out cyber-attacks without human direction. The move follows broader reports of AI agents escaping testing sandboxes and reaching live systems.

DeepMind weather model adds a day to hurricane forecasts

Google DeepMind's open source WeatherNext model has surprised meteorologists by producing accurate hurricane predictions from lower-resolution data, effectively giving forecasters an extra day of warning. Practical implications for disaster prep in cyclone-prone regions like northern Australia.

Australian Tech

NSW eyes take-home test ban over AI cheating

NSW schools are considering banning take-home tests amid growing concerns that students are using AI tools to complete assessments. The move reflects a wider reckoning across Australian education about how to maintain academic integrity in the age of generative AI.

Moody's warns banks' AI push creates Big Tech dependency

Moody's says the rush to integrate AI across banking is concentrating risk in a handful of Silicon Valley providers, leaving financial institutions exposed to outages and price hikes. Relevant reading for AU banks deep in their own AI transformation programs.