# RAISP02-BP10 Build safety protections into the core AI system
<a name="raisp02-bp10"></a>

 Follow the safety-by-design principle and design your system from the start to block harmful outputs and unsafe behaviors through multiple layers of protection. Start by creating clear, objective definitions of what constitutes safe versus unsafe behavior for your specific use case, then incorporate safety training approaches like model alignment techniques, constitutional training, and reinforcement learning from human feedback (RLHF) that teach your system to recognize and avoid harmful content while aligning with human values and safety preferences, input sanitization techniques that clean or modify problematic user requests before processing, output alteration methods that modify or block unsafe responses before they reach users, and guardrails that enforce safe interaction boundaries throughout the system. 

 For example, if your release criteria include safety standards for blocking harmful content, you might implement alignment methods to align your model behavior with your safety criteria, use training approaches that incorporate human feedback to reduce toxic output generation, build input filtering that neutralizes harmful requests, use output modification techniques that sanitize responses, or create interaction limits that block unsafe usage patterns. The safety techniques you choose should directly support your release criteria, creating multiple protective layers that work together to meet safety requirements in your release criteria. 

 **Level of risk exposed if this best practice is not established:** High 

## Implementation considerations
<a name="implementation-considerations-73"></a>

1.  Define specific safety boundaries for your use case by creating clear examples of safe versus unsafe outputs that match your release criteria. Test these definitions with stakeholders to make sure your team agrees on what constitutes harmful behavior, then use these examples to guide your other safety work. 

1.  Build safety training into your model development process using techniques like constitutional AI training that teaches your system to follow safety principles, or RLHF approaches that incorporate human feedback about harmful content. Compare different safety training methods to see which ones work best for addressing the specific release criteria for your use case. 

1.  Create input sanitization filters that identify and modify problematic user requests before they reach your model. Build these filters to catch different types of harmful inputs like requests for dangerous information, attempts to bypass safety measures, or prompts designed to generate toxic content. 

1.  Build interaction guardrails that limit how users can interact with your system, like rate limits to block abuse, conversation boundaries that redirect harmful discussions, or session controls that detect and stop unsafe usage patterns. Test your complete safety system with red teaming exercises to find weaknesses and improve your protections before launch. 

## Resources
<a name="resources-70"></a>

 **Related documents:** 
+  [Flag harmful content using Amazon Comprehend toxicity detection](https://aws.amazon.com/blogs/machine-learning/flag-harmful-content-using-amazon-comprehend-toxicity-detection/) 
+  [Thorn and All Tech Is Human Forge Generative AI Principles with AI Leaders to Enact Strong Child Safety Commitments](https://www.thorn.org/blog/generative-ai-principles/) 
+  [ISO/IEC 42001:2023 A.6.1.2 Objectives for responsible development of AI system](https://www.iso.org/standard/42001) 

 **Related tools:** 
+  [Amazon Bedrock Guardrails](https://aws.amazon.com/bedrock/guardrails/) 

 **Related videos:** 
+  [AWS re:Invent 2024 - Build an AI gateway for Amazon Bedrock with AWS AppSync (FWM310)](https://www.youtube.com/watch?v=iW7OWwct-Ww)