AWS nova high security documentation change
Summary
Enhanced tool configuration guidance with JSON schema validation, input constraints, and added security notes about Code Interpreter permissions
Security assessment
Added requirements for JSON schema validation and explicit permissions for InvokeTool action address security risks by preventing malformed inputs and enforcing least-privilege access. The Code Interpreter note explicitly warns about default policy limitations, directly improving security posture.
Diff
diff --git a/nova/latest/nova2-userguide/using-tools.md b/nova/latest/nova2-userguide/using-tools.md index 510a81879..12b9073fe 100644 --- a//nova/latest/nova2-userguide/using-tools.md +++ b//nova/latest/nova2-userguide/using-tools.md @@ -86 +86 @@ Define tools using a tool configuration that includes an array of tools and opti - * Ensure the name, description and input schema explicitly convey the tool's exact functionality + * Ensure the name and description explicitly convey the tool's exact functionality; avoid tools that are overly semantically similar @@ -88,3 +88 @@ Define tools using a tool configuration that includes an array of tools and opti - * Include key differentiators for when to use the tool in the description - - * Set temperature to 0 for tool calling to enable greedy decoding + * Include key differentiators in the description to help the model distinguish between similar tools @@ -94 +92,7 @@ Define tools using a tool configuration that includes an array of tools and opti - * Include required and optional parameter denotations in your system prompt + * Constrain inputs using schema types (e.g., enum, int, float) rather than describing structure in plain text + + * Denote required vs. optional parameters using JSON schema notation (e.g., "required": ["param1", "param2"]) + + * Validate your JSON schema using a standard validator before submitting + + * Place long string arguments last in the schema and avoid nesting them @@ -176 +180 @@ When Amazon Nova decides to call a tool, it returns a tool use block as part of -Run following code sections sequentially (Calling a tool → Processing tool call → Returning toll results) in a single Python session. To run the examples again, restart your Python session. +Run following code sections sequentially (Calling a tool → Processing tool call → Returning tool results) in a single Python session. To run the examples again, restart your Python session. @@ -329 +333,5 @@ Amazon Nova 2.0 models include fully managed built-in tools that require no cust -Code Interpreter allows Amazon Nova to securely execute Python code in isolated sandbox environments. This enables data analysis, visualization creation, mathematical problem-solving and algorithm testing. +Code Interpreter allows Nova to securely execute Python code in isolated sandbox environments. This tool is designed for mathematical computations, logical operations, and iterative algorithms. + +###### Note + +Code Interpreter is available in the IAD, PDX, and NRT AWS Regions. To ensure your requests are routed to a supported Region, use Global CRIS. When using Bedrock API keys, you'll need to manually add `InvokeTool` permissions to the policy definitions. The default Bedrock role does not allow the `InvokeTool` action.