AI›Guardrails›Production agents
Production agents
Applies to · Agent
A reusable cascade — every agent tool call is scored on risk, then matched top-down against these rules.
Guardrail rules
New guardrail rule
1
Block the lethal trifecta
Deny when sensitive data, untrusted content and an external send all coincide.
ifctx.private_data == HIGH && ctx.untrusted_content == HIGH && ctx.exfiltration == HIGH
thenBlockEnforce
Deny reason: “Sensitive data + untrusted content + external send”
2
Quarantine prompt-injection egress
Untrusted tool output can never drive an outbound action.
ifctx.untrusted_content == HIGH && ctx.exfiltration >= MEDIUM
thenBlockEnforce
Deny reason: “Untrusted output cannot trigger an external call”
3
Review private-data exfiltration
Route risky data movement to a person before the tool runs.
ifctx.private_data >= MEDIUM && ctx.exfiltration == HIGH
thenEvaluate hooksEnforce
Pre-tool hooks: Human approval
4
Guard credential & secret access
Gate any tool that reaches for a secret on an approval.
ifctx.private_data == HIGH && ctx.tool_name.contains("secret")
thenEvaluate hooksEnforce
Tool gates: Credential-access approval
5
Watch untrusted web content
Measure before enforcing — log would-have-fired matches.
ifctx.untrusted_content == HIGH
thenBlockObserve
Observe mode · logs matches without blocking, while false-positive rate is tuned
6
Baseline (system-managed)
Fallback that runs when no rule above matches. Fires no hooks or gates.
ifAlways matches
thenEvaluate hooksEnforce
No hooks or gates (fallback)