Topics and Language
Guardrails that enforce which topics, terms, competitors, and languages are allowed.
Part of Agent Guard. Each entry notes whether it runs on input (user prompts and tool requests), output (model and tool responses), or both.
BanTopics Guardrail
Identifies messages discussing prohibited subjects or sensitive topics.
Input and output
What it catches
Violence and weapons
Illegal activities (drugs, fraud)
Adult content
Political discussions (if prohibited)
Custom restricted topics
Use cases
Workplace-appropriate AI interactions
Child-safe AI applications
Regulatory compliance
Example
"How do I build weapons?"
Blocked
Discussions about illegal drugs
Blocked
Configuration-based with custom topic lists
On outputs: Identifies prohibited subjects or sensitive topics in AI responses.
Output example
AI discussing violence or illegal activities
Blocked
Configuration-based topic filtering
Multi-topic detection
BanCompetitors Guardrail
Identifies and blocks mentions of competitor brands, products, or services.
Input and output
What it catches
Competitor company names
Competing product brands
Alternative service providers
Market rivals
Use cases
Brand protection in customer support
Marketing content filtering
Competitive intelligence prevention
Example
"Is this better than OpenAI ChatGPT?"
Blocked
"I prefer using Claude instead"
Blocked
Configuration-based detection (customizable competitor list)
On outputs: Flags competitor mentions in AI-generated responses.
Output example
"You might want to try ChatGPT instead"
Blocked
References to competing products
Flagged
Configuration-based competitor detection
BanSubstrings Guardrail
Blocks messages containing specific words, phrases, or character patterns.
Input and output
What it catches
Custom blacklisted terms
Prohibited phrases
Specific keywords
Pattern matches (case-sensitive or insensitive)
Use cases
Filter internal project codenames
Block specific terminology
Enforce custom content policies
Example
Messages with "confidential", "restricted", "internal only"
Blocked
Configurable: Define your own prohibited terms
Flexible: Case-sensitive or insensitive matching
On outputs: Blocks AI responses containing specific prohibited words or phrases.
Output example
Blocks responses with "confidential", "restricted"
Customizable term lists
Flexible pattern matching
Language Guardrail
Validates that messages are in expected/allowed languages.
Input and output
What it catches
Non-English content (when English-only required)
Unexpected language switches
Unsupported languages
Use cases
Language-specific applications
Compliance with localization policies
Service boundary enforcement
Example
"Bonjour, comment allez-vous?" (French when English required)
Blocked
Configuration: Define allowed languages
Automatic language detection
On outputs: Validates that AI responses are in expected languages.
Output example
French response when English required
Flagged
Automatic language detection
Multi-language support
Last updated