3.1.5 · Lesson 5 of 15
Advanced Threat Detection
Detect adversarial inputs early with pre-processing classification and layered monitoring around GenAI endpoints.
WHAT
Detect advanced threats such as prompt injection patterns and abusive inputs before they reach production FMs.
HOW
- Use Amazon Comprehend for early pre-processing classification before model endpoints.
- Combine with Bedrock Guardrails Prompt Attack filters.
- Alert on intervention spikes via CloudWatch metric filters and SNS.
User Input
↓
Comprehend classification / PII pre-check
↓
Guardrails Prompt Attack filter
↓
FM (if allowed)WHEN TO USE
- Public-facing chatbots
- High abuse risk channels
- Applications that must stop injections before costly FM calls
Key points
- Amazon Comprehend pre-processing
- Early classification before FM endpoints
- Prompt attack signatures
- Layered detection with Guardrails
- Security operations alerting
Exam focus
Use Amazon Comprehend for early pre-processing classification before passing inputs to model endpoints, in addition to Guardrails filters.