# Security — community wiki

Zdroj: https://hermes-ai.cz/en/wiki/getting-started/security/
Ověřeno proti oficiální dokumentaci: 15. 10. 2026
Oficiální stránka: https://hermes-agent.nousresearch.com/docs/user-guide/security
Language: English. AI version — paste the whole text into your agent.

## Six safeguards
1. who may talk to the agent (account allowlist, DM pairing)
2. approval of dangerous commands
3. file protection (protected paths always blocked)
4. isolated runs in a container (sandbox)
5. project file scanning against injected instructions
6. secret filtering for what commands receive

## Command approval
- smart (default): asks for dangerous commands only
- YOLO (no asking): own server, no one else's data only
- hard floor: some commands are always refused, even without asking
- an approval is valid once; a command can be allowed permanently afterwards
  (hermes approvals suggest proposes what is worth allowlisting); /security shows the overview

## Unattended runs (scheduled jobs, external requests, single queries)
A dangerous command is simply denied - nobody being there to answer does not help.

## Files and secrets
- keys and passwords in the agent's settings, never in texts, skills or memory
- protected paths always blocked (agent settings, credentials)
- HERMES_WRITE_SAFE_ROOT narrows where the agent may write
- commands receive only a limited selection of environment values

## Tips
- no-asking mode is a bad idea on a machine with other people's data
- MCP connections from trusted sources only (they are third-party code)
- backups come before permissions