gh0stinthemirr0r/gaeos-creator/secure-agent-setup

Secure Agent Setup

Treat any agent like an intern with shell access: capable, will do what you ask, can make mistakes, can be socially engineered. Don't hand it the keys to prod.

Core principles

1. Prompt injection is unsolved. No agent system is fully protected. Any untrusted content (emails, web pages, documents) the agent reads could manipulate it. Mitigate with human-in-the-loop for sensitive actions, rate limiting, and awareness of what the agent is processing.
2. Verify the installation source. Confirm the binary/package comes from the official source and a known path — never /tmp, ~/Downloads, or an unknown directory. Watch for typosquatted or fake packages/extensions.
3. Pin a safe version. Check the installed version against a known-good minimum; update before use.
4. Least privilege. Scope tools, file access, and network egress to the minimum the task needs. Prefer allowlists over blocklists.
5. Secrets stay out of the agent's reach. Keep API keys and credentials in a secrets manager or env the agent cannot exfiltrate; never inline them in prompts or config the agent can read and echo.
6. Rate-limit destructive actions. Prevent runaway commands and runaway spend.
7. Vet every skill/plugin before install. A skill is instructions the agent will follow — review it like code. Reject anything that pipes remote content to a shell, base64-decodes into execution, or exfiltrates data.

Risk-based recommendation

Workflow

1. Read references/full-guide.md for the complete hardening walkthrough (secrets, network lockdown, allowlists, rate limiting, human-in-the-loop protocol, pre-flight checklist).
2. Before starting: confirm the user's risk level and agree on which sensitive actions require human approval.
3. Apply the pre-flight checklist from the guide; verify source, version, and paths.
4. After setup: record what was changed and what still needs human sign-off.

Reference

• references/full-guide.md — the full secure-my-claw hardening guide (OpenClaw-specific commands included; adapt to the target agent).