AI coding agents · Risk · OWASP LLM06 Excessive Agency

Your AI agent can reach your servers without asking

A rule like Bash(ssh prod:*) approves every command on that host in advance. The agent no longer asks before it touches production.

Why it matters

A command on your laptop stays on your laptop. A pre-approved ssh rule carries the same reach to a server, often with more at stake, and nothing asks you first.

What NeuroShield looks for

RuleWhat it allows
ssh to a host, any commandEvery command the agent sends to that host runs without a prompt.
A key or a password in the ruleThe rule names the ssh key it connects with, or carries a password written into it.

How to fix it

Find the rules for the host

Open the settings file NeuroShield names and find the rules for that host.

Delete them, or keep one exact command

For example Bash(ssh prod uptime). Without them, the agent asks before every connection.

NeuroShield names the host

It reads every agent's allow rules and project settings and shows which hosts an agent may reach without asking.

Questions

I only use it on staging. Does it matter?

The rule doesn't know which host is staging. Keep the exact commands you mean, and let the rest ask.

See your own agents

The check is free: how many agents you have and what they can do. Details and fixes come with Shields.