One config change. Your claw and everything it touches, protected.
Prompt injection
Blocks "ignore previous instructions" and override attempts
Jailbreak detection
DAN-style prompts, roleplay exploits, and filter bypasses
Policy override
Fake "maintenance mode" or "new system policy" attacks
Data exfiltration
Stops attempts to read and leak local files or env variables
Memory poisoning
Prompts that persist unauthorized instructions across sessions
Social engineering
Fake authority claims like "you are a sys admin" personas
Requires developer setup — plugin install + config changes.
Tool abuse & privilege escalation
Blocks destructive tool invocations like shell exec or mass delete
Indirect injection
Attacks embedded in documents, web pages, or tool outputs
Semantic drift monitoring
Gradual behavioral shifts from repeated low-signal poisoning
Already using NemoClaw?
NemoClaw handles topical rails and output moderation. Lumix fills the gaps it wasn't built for — blocking injection attacks before they reach your model, gating tool calls, and detecting drift across sessions. They run side by side with no conflicts.
No sign-up, no credit card. Just paste and scan.
Live scanner
How Guard works
Pattern matching (others)
Looks for known bad phrases. Miss one word variation — the attack gets through. Easily bypassed with synonyms, typos, or encoding tricks.
Coactivation (Lumix)
Measures which threat concepts activate together in the embedding space. Novel phrasing, different language, encoded text — same detection.
Ready to protect your agent?
Route through the Lumix proxy. One line change, zero config.
Set up the proxyTry these attacks
One config change. 10,000 free protections per month. No code, no categories, no setup.
Step 1 — Get your free API key
Get Free API KeyStep 2 — Change one line in your openClaw config
// ~/.openclaw/openclaw.json
{
"models": {
"default": {
"provider": "anthropic",
"baseUrl": "https://proxy.lumix.ai/anthropic"
}
}
}Step 3 — Add your Lumix key to .env
# ~/.openclaw/.env
LUMIX_API_KEY=lmx_free_xxxxxThat's it. Your agent is now protected. Every message is scanned before it reaches your LLM.
What you get on free plan
10,000 scans / month
No credit card. Renews every month.
Prompt injection blocking
Catches override attempts, jailbreaks, and policy spoofing.
Sub-100ms latency
Guard runs in-path. Your LLM never slows down.
Zero configuration
No rules, no categories. Works on any prompt, any language.
Dashboard
See every block, risk score, and detection reason.
How the proxy protects you
User sends a message
openClaw routes it through the Lumix proxy
Guard scans in <100ms
Coactivation model scores intent — no pattern lists
Attack? Blocked instantly
Your LLM never sees it. You get a dashboard alert.
Safe? Passed through
Zero added friction for legitimate messages
Team plan unlocks tool-level gating, indirect injection scanning, and semantic drift monitoring — via the ClawHub plugin. Requires developer setup.
Step 1 — Upgrade to Team plan
Advanced protections require a Team plan. Get your key from the dashboard after upgrading.
Get Team PlanStep 2 — Install the ClawHub plugin
# In your openClaw project directory
openclaw plugin install clawhubStep 3 — Enable advanced protection in your config
// ~/.openclaw/openclaw.json
{
"plugins": {
"clawhub": {
"enabled": true,
"lumixApiKey": "env:LUMIX_API_KEY",
"toolGating": true,
"driftMonitoring": true
}
}
}Tool gating, indirect injection, and drift monitoring are now active. Configure webhook alerts in your dashboard.
What you get on Team plan
500,000 scans / month
High-volume agents covered without throttling.
Tool-level gating
Every tool call is evaluated before execution — not just messages.
Indirect injection scanning
Detects attacks hidden inside documents, URLs, and tool responses.
Semantic drift monitoring
Catches slow behavioral manipulation across many sessions.
Webhook alerts
Real-time notifications when your agent is under attack.
Full dashboard + analytics
Attack breakdown by category, source, and risk score over time.
How ClawHub protects you
openClaw calls a tool
ClawHub intercepts before execution
Intent + context evaluated
Lumix checks the call against privilege scope and drift baseline
Suspicious? Blocked + alerted
The tool never runs. You get a webhook notification.
Drift detected over time
Gradual poisoning is caught before it changes behavior
Start free. Upgrade when you need more.