Agent Security Guard
From its catalog listing
Deterministic guard against prompt injection turning into actions. Results of web, search and browser tools reach the model wrapped as untrusted data; every planned tool call is checked against fixed action and sequence rules before it runs. After outside content entered a turn, a shell command, file write, install, config change or external send goes to Hermes's approval prompt; a secret is not sent out through web, e-mail, upload or push tools; the agent does not rewrite its own guard or skills on its own initiative. No model decides; every decision is recorded in a hash-chained audit trail.
Good to know before installing
no network calls, no shell commands, no background process, no credentials, no Python dependencies. Reads its policy from /etc/agent-security-guard/guard.yaml or guard.yaml in the Hermes home; loads its rule engine from /usr/local/lib/agent-security-guard/src or agent-security-guard/src in the Hermes home ahead of the pinned copy when an operator installed one there; for file tools it resolves the named path (realpath, stat) without reading the file; writes its audit trail (mode 0600) to $XDG_STATE_HOME/agent-security-guard. Sees tool calls, not what runs inside a shell command; not a sandbox.
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 agent-security-guard - Said no? Turn it on later:
hermes plugins enable agent-security-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
- xMannixx (Community)
- Needs
- No API keys declared
- Adds hooks
pre_llm_callpre_tool_calltransform_tool_result- Needs Hermes
- 0.21.6 or newer
- Version
- 0.4.4
- In the catalog
- 2026-11-10
Known issues (from its catalog listing)
- Sees tool calls, not what happens inside a shell command; the approval prompt after outside content is the protection there.
- Messaging tools (send_message) are audited, not stopped, also when the text holds a secret.
- Browser actions (browser_exec) and input to a running process (process_manage) are not told apart from reading; third-party tools it does not know stay free after outside content unless declared in tool_tiers or mode is strict.
- One audit trail per OS user; Hermes profiles each read their own guard.yaml but share the trail.
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