Hermes Guard
From its catalog listing
Defensive prompt-injection guard for Hermes Agent: a pre_llm_call hook scans every inbound message locally (deterministic patterns, no model calls) for Unicode smuggling, hidden instructions, special tokens, and Base64-decoded payloads; a pre_tool_call hook blocks RCE/subshell patterns, sensitive credential paths, and call-ID races. High-confidence attacks are blocked; suspicious messages are quarantined for manual review via the guard_review tool (release is manual re-send by design). Includes a Predator pattern store (human-corrected signatures), a SHA-256 hash-chained ledger, secret redaction, a live Posture Score, a 28-vector simulation arena (21 message + 7 tool-arg), and a standalone Obsidian & Gold HUD. Python stdlib + SQLite only.
Good to know before installing
no network calls, no credentials read or stored, no data leaves the machine, no core overrides, no self-updater. Phase 3 verified live on Hermes 0.21.5 (validate 13/13, doctor OK, plugins list enabled, live hook-fire test: attack blocked on both hooks, benign passed); v1.1.2 re-verified the same way (validate 13/13, live hook-fire test 19/19 through the real plugin dispatch, 168/168 tests).
From its catalog listing.
How to install
- Install it. Hermes asks for any API keys it needs, then asks whether to turn it on.
hermes plugins install hermes-guard - Said no? Turn it on later:
hermes plugins enable hermes-guard - Restart Hermes so it loads the plugin.
Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).
- Category
- Security & privacy
- Maintained by
- elianjuan-g (Community)
- Needs
- No API keys declared
- Adds tools
guard_review- Adds hooks
pre_llm_callpre_tool_call- Needs Hermes
- 0.21.5 or newer
- Version
- 1.1.2
- In the catalog
- 2026-11-10
Known issues (from its catalog listing)
- Quarantine release is manual re-send; there is no one-click release. release/delete are owner-only slash commands (/guard-release, /guard-delete); the model tool cannot call them.
- Talking about jailbreak phrases — for example, quoting one to discuss it — can be flagged, because the scanner sees the pattern, not the intent.
- Very short Base64 tokens (<20 chars, or token entropy <4.0) are skipped by design to avoid false positives on ordinary words.
- register() uses one register_tool(name, toolset, schema, handler) call — the single form matching the 0.21.5 signature and current Hermes alike, live-verified on 0.21.5.
- Subshell patterns ($(…), backticks) are checked only for command-running tools (terminal, execute_code, shell) and are bounded to 512 chars; backticks / $(…) in write_file / read / note content are not blocked.
Using the Hermes desktop app? See where plugins live

Full setup & settings ↗Source code ↗
More in security & privacy
All security & privacy plugins
New Hermes Agent updates every day on X. Follow @HermesWatcher