The Agent Skills Directory

[SAFE]: The skill functions as a security auditing tool. It uses official cloud SDKs to verify configuration against established security benchmarks (NIST SP 800-207) and does not perform any destructive actions or unauthorized data transmission.
[INDIRECT_PROMPT_INJECTION]: The assessment script (scripts/agent.py) ingests metadata from AWS APIs, such as IAM user names and security group descriptions. This content is technically external/untrusted data. However, the script only performs read-only diagnostic reporting and does not execute the ingested data or transmit it to external services.
Ingestion points: iam.list_users(), ec2.describe_instances(), and s3.list_buckets() in scripts/agent.py.
Boundary markers: None present in the script output.
Capability inventory: Limited to read-only boto3 calls for auditing purposes; results are printed to stdout or saved to a local JSON file.
Sanitization: None, as the data is used for reporting only and not for further command generation or interpolation into prompts.
[CREDENTIALS_UNSAFE]: The skill instructions in SKILL.md contain placeholders for OIDC client secrets (verified-access-client-secret). These are clearly identified as placeholders for user configuration and do not represent hardcoded credentials.

implementing-zero-trust-in-cloud