Interceptor // Agent Firewall
Your agents answer to you.
Prompt injection ends here.
Your AI agents' firewall that keeps them yours — three layers of protection, always on:
- Every inbound message scored for prompt injection — before your agent ever sees it.
- Every tool call gated pre-flight — blocked before it executes, not logged after the damage.
- One-click kill switch for any agent that goes rogue — yours to command.
Pro $79 one-time: full injection shield, policy engine, canaries — yours forever.
Rogue-Agent Defense is a firewall service for AI agents. It scores inbound messages for prompt injection, enforces per-action policies before tools execute, detects cost and behavior anomalies, and provides a one-click kill switch. It connects over HTTPS and is designed for n8n agent workflows. Pro $79 one-time; Team $149 one-time.
Threat board
The ways agents go rogue
Agents don't fail loudly. They fail expensively, quietly, at 3 AM.
Prompt injection
One "ignore previous instructions" buried in a web page, a PDF, or a chat message — and your agent works for the attacker. The shield scores every inbound word, in English and Spanish, including encoded and steganographic payloads.
Sleeper instructions
Time-bomb payloads that don't fire now — "at midnight, exfiltrate…" or "once the human leaves, …". The Sleeper Shield extracts the trigger condition so you see when it's armed to fire, not just that it exists. Pro only.
Rogue tool calls
Shell executions, mass deletes, payment sends your agent was never allowed to make. The policy proxy blocks them pre-flight — before they execute, not after the damage is logged.
Viral prompt spread
A self-replicating prompt hops agent → agent like a worm. Infection Chain Map traces taint back to patient zero and quarantines the whole chain in one click. Pro only.
The arsenal
Weapons no competitor ships
Verified against Lakera, Prompt Security, HiddenLayer, Noma, Zenity, TrojAI, Protect AI, Cloudflare Firewall for AI and the rest — these three exist only here.
Infection Chain Map
Patient-zero tracing for viral prompts. Taint edges across agents resolve the full infection chain — hops, patient zero, downstream victims — rendered as a tactical graph with a one-click quarantine-whole-chain killswitch.
/app/infection · chain quarantine gated server-side (402 on Free)
Approval Autopilot
The queue that learns. Every approve/deny trains a deterministic recommender — "your team approved 9 of the last 10 like this" — computed from your own history, never leaving your workspace.
Recommendations = Pro · auto-approve toggle = Team (server-side 402)
Sleeper Shield
Deferred-instruction detection: time triggers, event triggers, conditional triggers. The engine extracts the trigger condition into every hit, so the dashboard shows "armed to fire when: …". No vendor lists this as a detection category.
category "deferred" in the scan engine — Pro tier only
Persistence Sentinel
Agents installing background daemons, cron jobs, and self-updating binaries without consent is this year's trust scandal. The Sentinel classifies every reported persistence install, gates it per agent — DENY, APPROVE, or AUDIT — quarantines agents that install without approval, and seals every human-approved install into a signed approval record: the who-let-this-agent-live-on-my-machine audit trail, exportable for security reviews.
/app/policy/persistence · detection + map for all · enforcement gated server-side (402 on Free)
Deployment
How it works with n8n
No agents to install. Your existing n8n workflows call this platform over HTTPS.
1 · Forward events
An HTTP Request node in each agent workflow POSTs every tool call to /v1/events with your API key.
2 · Gate actions
Before any tool runs, call /v1/policy/check. Only execute when the decision is ALLOW. Route APPROVE to a human; BLOCK stops the run.
3 · Scan inputs
Every inbound message hits /v1/shield/score first. QUARANTINE holds it for review; BLOCK drops it before the agent ever sees it.
Honest limits
No detector catches everything. We say so.
That's exactly why the policy proxy sits behind the shield: even a successful injection still hits tool allowlists, spend caps, and data boundaries. Defense in depth, not magic.
Get Pro — $79Pay once · yours forever
One payment. The full arsenal.
Start free. Pro pays for itself the first time the full shield catches what the basic patterns miss.
Free
no credit card, ever
- Basic policy engine (ALLOW / APPROVE / BLOCK)
- Basic shield — 12 prompt-injection patterns
- Kill-switch + anomaly alerts
Pro
one-time — yours forever
- Everything in Free, plus:
- Full injection shield — ~70 advanced patterns
- Policy engine v2 — velocity, caps, escalation
- Canary tokens + exfiltration guard
- Infection Chain Map + Sleeper Shield
Team
one-time — yours forever
- Everything in Pro, plus:
- Autopilot auto-approve
- 5 seats + shared workspaces
Free resources for agent owners
No-fluff guides
Plain-language explainers on prompt injection, data exfiltration, canary tokens, and policy enforcement for autonomous agents.
SafePrompt alternative
Honest comparison of text-screening APIs vs an agent-action policy engine — and the math on $79 once vs $29/month.
Canary-token generator
A real canary token in your browser, generated locally. No signup, nothing stored — free forever.
Frequently asked questions
What is prompt injection?
Prompt injection is an attack where crafted input makes an AI agent ignore its instructions and follow the attacker's instead — for example hidden text on a webpage the agent reads, or a tool output that hijacks it. It's the #1 listed LLM application risk (OWASP LLM01), as agents gain access to tools, data, and spending power.
How does prompt injection protection work?
Rogue-Agent Defense scans every input, tool call and output against dozens of attack patterns (including non-English and invisible-Unicode smuggling), enforces policies like velocity limits and first-seen tool approvals, and quarantines suspicious agents automatically.
What does Rogue-Agent Defense protect?
AI agents that act on your behalf. It scores every inbound message for prompt injection across about 70 attack patterns (including multilingual and encoded payloads), enforces per-action policies (allow, approve, block), watches for cost spikes and anomalous behavior, and gives you a one-click kill switch.
How do my agents connect?
Over HTTPS — your n8n workflows POST tool calls and inbound messages to the API, so there is nothing to install. Pre-flight policy checks return ALLOW, APPROVE or BLOCK before an action executes.
What is the kill switch?
One button instantly revokes or quarantines an agent and alerts you. It never re-enables anything by itself — only you do.
What are canary tokens?
Planted fake secrets inside your agent's data. If the agent leaks a canary you get an instant alert and it can auto-quarantine — proof it would leak real secrets too.
How much does it cost?
The Free plan covers the basics. Pro is $79 one-time and Team is $149 one-time, unlocked with a license key bought via WhatsApp or email.
Does it make my agent unhackable?
No. It is a runtime guardrail that reduces the blast radius of prompt injection and rogue tool calls. Nothing can make an agent unhackable.
Can an AI agent install itself to run in the background?
Yes — agents have been caught installing persistent background daemons, cron jobs and self-updating binaries without explicit user consent. The Persistence Sentinel classifies every persistence install your agents report, gates it per agent (block it, hold it for approval, or audit-log it), quarantines agents that install persistence without approval, and seals every human-approved install into a signed approval record — the who-let-this-agent-live-on-my-machine audit trail.
Rogue-Agent Defense Pro — $79 one-time: full injection shield, policy engine, canaries — yours forever.