The Agent Skills Directory

Management of Untrusted Data: The skill provides instructions for building RAG (Retrieval-Augmented Generation) systems that ingest data from external sources such as web crawlers, S3, and third-party SaaS connectors. Processing untrusted data is a consideration for indirect prompt injection. To mitigate this, the skill includes detailed guidance on implementing Bedrock Guardrails, using isolation blocks like guardContent to separate untrusted context, and validating all agent-generated parameters.
Command Execution via Automation Scripts: The skill generates and executes local Python scripts (e.g., provision_managed_kb.py and setup_payments.py) to automate the provisioning of AWS resources. These scripts are created from internal templates and configured with user-provided data. The workflow enforces security through mandatory review phases, dry-run previews, and explicit user approval before any commands are executed.
Secure Credential Handling: Guidance for configuring API keys and OAuth tokens emphasizes security by instructing users to set environment variables rather than hardcoding values. Automation scripts are designed to read these credentials from the environment, and the skill provides clear warnings about the risks of exposing secrets in shell history or code.
Least-Privilege and Access Control: The skill provides templates for scoped IAM policies and resource-based permissions. It emphasizes restricting access to specific ARNs and includes condition keys like aws:SourceAccount and aws:SourceArn to protect against confused deputy attacks during resource provisioning.
Official External Tooling: The instructions facilitate the installation of verified vendor tools, such as the AgentCore CLI and standard SDKs (e.g., boto3), from official registries. These resources originate from the vendor's verified organization.

amazon-bedrock