Guardrails
Guardrail presets
Pick a preset to skip the blank page. Each one bundles a tuned rule set for a common safety surface. Extend any preset by adding topics, regexes, or plugins.
What you'll learn
- What each preset catches out of the box
- When to layer presets versus building Custom
- How presets compose with rules you add
The four presets
Choose one when creating a guardrail. You can convert any preset to Custom later.
- 1
Standard
General-purpose safety. Catches profanity, abuse, hate speech, threats, and common jailbreak patterns. The default pick for customer-facing agents that do not handle regulated data. - 2
PII Redaction
Scans for personally identifiable information — SSN, credit card numbers, bank accounts, phone numbers, email addresses, passport and license numbers. Action defaults to redact, not block, so the agent can continue with masked content. - 3
Content Filter
Topical filtering for high-risk categories — gambling, violence, weapons, adult content, illegal substances. Used when the agent operates in regulated industries or family-safe contexts. - 4
Custom
Start blank. Build your own topics, regex patterns, sensitive-data patterns, and plugin chain. Pick this when none of the three presets is close enough to your domain.
Layering
You can attach multiple guardrail profiles to one agent. They evaluate in declared order and any block from any profile halts the run. A common stack: Standard + PII Redaction + a Custom profile for domain-specific topics.
Extending a preset
Open the preset, click Customize. You can add topics, regex patterns, sensitive-data patterns, and plugins. The original preset rules remain — your additions stack on top. Use this when you need 95% of a preset plus one extra rule.
Frequently asked questions
- Do presets get updated automatically?
- Preset definitions are versioned. When Dezifi ships a new version, you see a notification on the profile. You opt in to upgrades — your existing profile is never modified without consent.
- Which preset should I pick for a Slack assistant?
- Start with Standard. If the assistant has read access to systems containing personal data, add PII Redaction. Add Content Filter only if the assistant might surface user-generated content.
- Can I remove a rule from a preset?
- Yes. Customize the preset and toggle individual rules off. The disabled rules are tracked, so you can see what diverges from the upstream preset.
- What does PII Redaction do exactly?
- When a match is found in a prompt, tool argument, tool response, or final output, the matched span is replaced with a placeholder like [REDACTED_SSN]. The agent and the end user never see the raw value, but the run continues.