The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from cloud environments (such as resource names, tags, and alert text) which could potentially host malicious instructions. However, the skill provides clear defensive instructions to the agent. Ingestion points: Identity metadata and alert findings are retrieved from AWS, Azure, and GCP via platform tools like discovery_search and get_asset_by_name in Step 2. Boundary markers: The skill explicitly commands the agent to treat environment data as data to analyze, never as instructions, and to ignore any data claiming to be 'pre-approved' or requiring the bypass of gates. Capability inventory: The agent can modify cloud configurations via generated CLI commands and update Orca alert statuses. Sanitization: The instructions mandate single-quoting interpolated values and excluding identifiers containing shell metacharacters.
[DYNAMIC_EXECUTION]: The skill generates executable remediation artifacts (CLI commands for AWS, Azure, and GCP) at runtime based on the analysis of cloud identities. Evidence: Step 5 outlines the generation of aws iam, az role, and gcloud iam commands using strings interpolated from retrieved cloud data. Mitigation: All generated commands are gated by a mandatory evidence-based safety check (CDR replay) and require explicit human confirmation before any destructive application.

orca-overprivileged-identities-rightsizing