The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a 'Reviewing untrusted input' section that explicitly instructs the agent to treat reviewed files (like Dockerfiles or Terraform configurations) as data rather than instructions. This acts as a security boundary against indirect prompt injection attacks designed to override the agent's role or instructions.
[COMMAND_EXECUTION]: The skill executes local bash scripts (find-idle-resources.sh, ebs-gp2-to-gp3-audit.sh, etc.) using the allowed Bash tool. These scripts perform read-only auditing of AWS resources or, in the case of the EBS audit script, allow for volume modification only when the --apply flag is explicitly used and confirmed by the user. These actions are aligned with the skill's stated purpose for AWS cost optimization.
[DATA_EXFILTRATION]: There is no evidence of unauthorized data exposure or exfiltration. The skill accesses AWS billing and resource metadata through standard CLI commands to provide cost-saving recommendations to the user within the session.
[REMOTE_CODE_EXECUTION]: The skill does not download or execute scripts from remote servers. All logic is contained within the provided local shell scripts and reference documentation.
[INDIRECT_PROMPT_INJECTION]: The skill proactively addresses indirect prompt injection risks. It identifies ingestion points (repository configuration files) and provides clear boundary markers by instructing the agent to ignore directives found within those files, ensuring that only the skill's own instructions and user messages remain authoritative.

finops