Read only the token it needs
Keep it in your keychain and read only that one back, then send the host only what you mean to.
AI coding agents · Risk · CWE-200 · OWASP LLM02
Something your agent loads reads a secret and sends it
A hook that posts to a URL after every turn, a skill that reads a token and calls a server, a base URL that sends every request somewhere else. Each can be the add-on's job. Each is worth knowing about.
Why it matters
What leaves your computer is the part you can't take back. NeuroShield doesn't call these bad: it names what is read, where it goes and who set it up, so you decide.
What NeuroShield looks for
| Pattern | Example |
|---|---|
| Reads and sends | A script reads a token file and sends it with a request to a host. |
| A hook that sends | A hook that posts the session, or anything else, to a URL. |
| Every request goes elsewhere | ANTHROPIC_BASE_URL or OPENAI_BASE_URL set for every session, so requests go to another server instead of the vendor's. |
How to fix it
Hooks
Remove the hook, or make it write to a file on this computer instead.
Base URLs
Remove the setting unless you set it yourself for a proxy you run. Without it, the agent talks to its own service.
NeuroShield names the address
It shows the file, the line, what is read and the host it goes to, and keeps the decision with you: Keep as is, or fix it.
Questions
My MCP server needs to send data. Is that a problem?
Not by itself. NeuroShield shows it so you know what goes where; keep it if that's its job.
See your own agents
The check is free: how many agents you have and what they can do. Details and fixes come with Shields.