What’s new: For JetBrains users with a Copilot plan: cloud tasks run on the Copilot subscription instead of JetBrains AI credits, each in a fresh isolated virtual machine.

What’s new: For teams running code scanning: it adds an opt-in threat model for LLM-generated data and fewer false positives in several queries. Custom GitHub Actions queries may need updates.

What’s new: For teams sharing agent instructions across repos: they can now be kept current like any other dependency. It covers version updates, not security updates.

What’s new: For people building computer-use agents that run on a device. The authors report 66.0% on the OSWorld-Verified benchmark, which they call best at its size.

What’s new: For robotics tinkerers who want to run a local model on real hardware. Its authors claim Qwen3.8-Flash-Next with it beat a fine-tuned robot model on real tasks.