Safety
"Never do X"
—and it actually never does
Define rules in plain English or code. Your AI physically cannot violate them—not through jailbreaks, not through edge cases, not ever.
Guardrails
Interactive preview
SafetySafety
Documentation →Rules that can't be broken
Not guidelines. Not suggestions. Architecture-level enforcement that can't be bypassed.
Plain English or code
'Never discuss competitors' or complex regex—write rules however makes sense for you.
Violations get handled
Block the response. Modify it. Flag for review. You decide what happens when rules are tested.
Rule builder
Create guardrails with natural language or code. Test before deploying.
- Natural language rules
- Code-based rules
- Testing sandbox
CODE
empress.guardrails({
query: "show me flagged accounts",
options: {
limit: 10,
timeRange: "7d"
}
})Enforcement engine
Rules evaluated in real-time. Violations handled per your configuration.
- Real-time checking
- Multiple actions
- Exception handling
EXECUTION FLOW
1
Input received2
Context retrieved3
LLM inference4
Tool execution5
Output generatedHow it works
1
Define
Create guardrail rules
2
Test
Validate rules in sandbox
3
Deploy
Enforce rules in production
Similar in Safety
All apps →Set guardrails
Boundaries that hold.
Request beta access