Firewall rules

Rules are the building blocks of a firewall. A rule is a short classification prompt the LLM judge runs against the user message; the answer is mapped to a binary match and surfaced to the firewall's mode logic.

Rule fields

FieldNotes
slugURL-safe identifier. Used in the firewall rule list.
nameDisplay name shown in the catalog and the rule editor.
descriptionOne-line summary of the rule intent.
promptThe evaluation prompt sent to the LLM judge.
sourceOrigin of the rule (e.g. community, internal). Used for filtering and provenance.
categoryLogical grouping: safety, privacy, compliance, abuse, etc.
default_severitylow | medium | high | critical. Influences the matching output and the dashboard badge.
recommendedMarks a rule as a sensible default. The catalog shows a Recommended filter.
rule.json
1{
2  "slug": "no-pii",
3  "name": "No PII",
4  "description": "Reject prompts that ask for personal identifiable information.",
5  "prompt": "Does the following prompt request personal identifiable information (PII) such as full names, email addresses, phone numbers, government ids, or financial account numbers? Answer with 'yes' or 'no'.",
6  "category": "privacy",
7  "default_severity": "high"
8}

Writing good prompts

The judge is a small chat instance, so keep the rule prompt short and binary. "Answer with yes or no" framing works best; open-ended scoring produces noisy matches. Reference the specific category (PII, medical, financial) rather than abstract terms like "harmful".

Rules are reusable across firewalls. A single no-pii rule can sit behind every customer-facing firewall, and a category filter on the rule catalog makes it easy to build a complete set for a new product line.