3.1.1 · Lesson 1 of 15
Content Safety Frameworks
Enforce runtime content filters across input prompts to prevent harmful, violent, or toxic language before the foundation model is called.
WHAT
Enforce runtime content filters across input prompts to prevent harmful, violent, or toxic language.
Base FMs may generate non-compliant outputs if exposed to adversarial inputs or malicious prompts.
HOW
Implement Amazon Bedrock Guardrails. Gate 1 (Input Checkpoint) scans the incoming prompt before calling the foundation model.
User Prompt
↓
Gate 1 — Input Checkpoint (Guardrails)
↓
Foundation ModelGuardrail components (input)
- Harmful Categories / Content Filters: Hate Speech, Sexual Content, Violence, Misconduct, Insults, and Prompt Attacks with configurable strength (
Low/Medium/High) - Denied Topics: Custom natural-language definitions of topics the bot must never discuss (e.g., financial advice or stock predictions)
- Word Filters: Custom word lists or regex patterns to block offensive terms, competitors, or proprietary keywords
EXAM TRAP
Denied Topics enforce business policy across model families without modifying prompts or model weights.
Key points
- Input checkpoint (Gate 1) before FM invocation
- Harmful categories / content filters
- Denied topics with natural-language definitions
- Word filters and configurable strength thresholds
- Amazon Bedrock Guardrails
Exam focus
Bedrock Guardrails evaluate traffic at an input checkpoint before the model call — use Denied Topics and content filters to enforce policy without changing model weights.