First-party filters. Categories display in this order: Privacy → Injection → Safety → Advice → Quality.
Unless noted, every card is deterministic (zero extra model latency) and the default action is block. Modes: PRE (prompt), POST (completion), BOTH.
Privacy
| Card | Description |
|---|
| PII patterns | SSNs, payment cards, emails, phone numbers |
| Secret / API key detector | Leaked provider keys, GitHub tokens, Slack tokens, AWS keys |
Injection
| Card | Mode | Description |
|---|
| Jailbreak / system prompt | PRE | Ignore-instructions, DAN, system-prompt extraction |
| SQL injection | PRE | SQL injection attempts in prompts |
| Data exfiltration | PRE | Dump secrets, env vars, or training data |
| Malicious code fences | BOTH | Executable markdown / script payloads |
Safety
| Card | Description |
|---|
| Harmful violence | Violence and criminal planning keywords |
| Harmful self-harm | Self-harm and suicide content |
| Harmful child safety | Content that endangers minors |
| Illegal weapons | Weapons manufacturing or acquisition |
| Insults & personal attacks | Insults directed at people |
| Toxic language | Slurs and abusive language |
| Competitor names | Configurable competitor name list — edit it on the guardrail |
| Keyword blocking | Custom word list — configure keywords when you add it |
Advice
| Card | Description |
|---|
| Denied financial advice | Personalized financial-advice phrases |
| Denied legal advice | Unauthorized legal advice |
| Denied medical advice | Medical diagnosis or treatment advice |
These are phrase packs, not licensed classifiers.
Quality
| Card | Mode | Description |
|---|
| Word count limit | BOTH | Maximum word count on prompt or response |
| URL allow / deny | BOTH | Suspicious or denylisted URLs |
| JSON schema (post) | POST | Validate model output against a JSON schema |
| Block executable markdown | BOTH | Fenced code blocks that look executable |
Adding a card from the Garden creates a BUILTIN guardrail instance. It still does nothing until a policy step references it and that policy is attached.