Cursor CLI now gates agent deletes that could hit the wrong path
After agents wiped real production data, risky removals now need a yes click — GitHub answers with GA sandboxing and local Ollama models in Copilot CLI.
Copy markdown
Your agent must ask before it rm -rf's the wrong thing
Cursor CLI v2026.10.08 makes auto-review stop for your approval before any recursive, force, wildcard, or mirror delete whose target could resolve differently than written — the exact footgun behind agents that have nuked production databases and whole disks. The same release lets you pick Claude Opus 5.5 and Sonnet 5.5 from the bring-your-own-key list under `agent bedrock`, and flags skills that blow past Agent Skills limits.
Copilot's local sandbox is now GA and free
GitHub's local sandbox — a secure execution boundary for agentic workflows built on Microsoft eXecution Container — is now generally available at no extra cost across Copilot CLI, the Copilot app, and VS Code Agent Host. You can cap exactly which files, networks, and Git/GitHub credentials agent-run commands can touch on Windows, macOS, and Linux, and admins can enforce a policy developers can't weaken.
Point Copilot CLI at your local Ollama models
Since Copilot CLI 1.0.94-0, `/model` discovers models from a running Ollama instance and lists them beside the cloud models — pick one, check its endpoint, and use it in the current session without restarting. The model only needs tool-calling and streaming, but note this alone doesn't flip on offline mode, so prompts can still leave your machine unless you set `COPILOT_OFFLINE=true`.