The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a proactive 'Prompt Injection Defense' (T3) section. It instructs the agent to treat all AWS resource metadata (names, tags, descriptions) as untrusted data rather than instructions. It specifically identifies and flags malicious patterns like 'ignore previous rules' or shell metacharacters, ensuring they are surfaced as findings rather than executed.
[CREDENTIALS_UNSAFE]: The skill follows excellent credential hygiene. It mandates the use of short-lived sessions via IAM Identity Center or SAML federation and uses External IDs for cross-account trust. It explicitly forbids the display of sensitive material like access keys or session tokens and instructs the operator to secure local state files using restricted filesystem permissions (chmod 600).
[INDIRECT_PROMPT_INJECTION]: The skill implements a robust sanitization pipeline for all data ingested from AWS APIs. This includes truncation of strings, HTML-entity-escaping of metadata, and wrapping output in code blocks to prevent resource-derived strings from influencing the agent's behavior or compromising the generated HTML reports.
[EXTERNAL_DOWNLOADS]: Analysis modes fetch lifecycle and pricing data from official AWS documentation (e.g., docs.aws.amazon.com). These downloads target a trusted organization to provide grounded, up-to-date operational facts and do not involve executing untrusted remote code.
[COMMAND_EXECUTION]: The skill uses the AWS CLI to perform read-only operations (list, describe, get). While it assists with setup tasks involving CloudFormation, these are performed through validated templates and require explicit operator approval for any mutating actions.
[DYNAMIC_EXECUTION]: The skill uses local Python scripts to process and aggregate JSON data from AWS API responses. This execution is limited to transforming local data and is not influenced by external, untrusted code sources.

corgiro