Safety features

As it is most likely dangerous if Neuro was given free reign, NeuroPilot has built-in safety features enabled by default. This page will go over them, what they do, and the consequences of disabling them.

Tip

This page basically reiterates information already put into other pages for the sake of unifying safety features into one page.

Danger

Enabling direct terminal access will disable all safety features inherently. This is mostly due to being able to modify .vscode/settings.json and being able to execute whatever CLI commands she wants.

It is always better to pre-define tasks for Neuro to run than giving her terminal access.

Incompleteness

An unintentional safety feature within NeuroPilot lies within its incompleteness: at the moment, it does not fully recreate Visual Studio Code 1:1. Until that is achieved, the incompleteness factor will prevent Neuro from performing complete takeover.

File access restrictions (Neuro-safe path checks)

NeuroPilot has a set of checks (dubbed Neuro-safe path checks), which checks if a path specified in inputs is considered “Neuro-safe”.

By default, the following paths are considered Neuro-unsafe paths (they will be checked against for path-related actions and if present, NeuroPilot will refuse to execute the action):

• Files and folders starting with a . (such as .vscode/** or .env)
• Environmental variables
• Parent directories
• Home directory (~)
• The workspace folder itself

You can add additional Neuro-unsafe paths using the Include Patterns setting.

You can also disable the default Neuro-unsafe path checks using the Allow Unsafe Paths setting. If you do so, Neuro will be able to access potentially unsafe file paths, such as .env (which is where you’d store your secrets, such as API keys) and .vscode/settings.json (which could allow her to edit NeuroPilot’s settings, potentially giving herself Autopilot-level permissions and disabling safeguards).

Opt-in task action registration

NeuroPilot only registers tasks whose detail description start with [Neuro] (case-insensitive). Each action will have the Neuro tag substringed from their detail description before registration (and trailing whitespaces stripped).

You can see an example in the playground directory of the NeuroPilot repo.

You can allow Neuro to use ALL defined tasks by enabling the Allow Running All Tasks setting. This will allow Neuro to run any tasks defined in .vscode/tasks.json. Tasks that have been marked with the Neuro tag will still be substringed to anonymise what is normally registered anyways. The impact of enabling this setting vastly depends on the tasks defined in there.

Explicit shell definition

More info: Settings/Terminals

In order to run shells, they must be explicitly defined in .vscode/settings.json with their name, path and args to be passed in (if any).

You can find out how to define them by reading the link above. You cannot disable this safety feature.

Permissions

The action permissions system is the biggest safety feature within NeuroPilot. This defines how Neuro can use certain actions - either allowing her to do it herself, require approval, or prevent her from doing it. This system is tightly integrated into NeuroPilot, and disabling it can mean two things:

• Setting action permissions to Autopilot: Neuro will be able to execute that permission whenever she wants without needing approval. You may want to look at the permissions reference to see what each permission group can do.
• Disabling all permissions by setting them to ‘Off’: Neuro will no longer be able to run any actions unsupervised. You can still ask her questions via Copilot Chat without any problems.

You can learn about the way NeuroPilot handles action permissions on the RCE page.

You cannot disable the RCE framework itself as that would effectively mean every Autopilot-able permission will no longer function. However, what can be construed as effectively disabling it is setting every permission to Autopilot mode. This will allow Neuro to do any and all actions whenever she wants without pre-execution human oversight.

Disable All Permissions command

As a panic button, you can activate the Disable All Permissions command. This will immediately shut off access to all permissions, even requesting cookies, and will also forcibly change certain non-action-permission settings.

You cannot fully disable this command. However, you can choose not to bind it to a keyboard shotcut. This will mean that you cannot instantly disable permissions for Neuro if she is misbehaving.