AI Safety, Security, and Governance
Input and Output Safety Controls
Apply layered defenses against harmful content, prompt injection, jailbreaks, hallucinations, data exfiltration, and unsafe tool use before, during, and after model inference.
Concepts
- Treat user input and retrieved documents as untrusted data, not authoritative instructions.
- Combine managed guardrails, deterministic validators, grounded retrieval, authorization, and monitoring for defense in depth.
- Structured outputs narrow downstream ambiguity but must still be parsed and validated outside the model.
- Adversarial test suites should cover indirect injection, obfuscation, excessive agency, and sensitive-data leakage.
Exam tips
- Bedrock Guardrails can filter denied topics, harmful content, and sensitive information.
- Guardrails complement rather than replace application authorization and tool validation.
- Grounding and citations help accuracy, but retrieved content can itself contain malicious instructions.