The Agent Skills Directory

[DATA_EXPOSURE]: No hardcoded credentials, API keys, or access to sensitive local file paths (e.g., .ssh, .aws) were found.
[EXTERNAL_DOWNLOADS]: The skill does not perform any external network requests or download third-party scripts.
[COMMAND_EXECUTION]: No dangerous shell commands, privilege escalation (sudo), or persistence mechanisms are present. The skill uses standard sub-agent spawning for its verification logic.
[PROMPT_INJECTION]: The skill does not contain instructions to bypass safety guidelines or override agent behavior. The adversarial evaluation mentioned is a quality control technique to find errors in generated output, not a security bypass.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data (generated output and task specifications) during the review phase.
Ingestion points: Untrusted data enters the reviewer context via the {output} and {task_spec} variables in the REVIEWER_PROMPT (SKILL.md).
Boundary markers: Uses Markdown headers as delimiters for input sections.
Capability inventory: The skill spawns independent sub-agents using the platform's Agent tool to perform reviews and a fix_agent to modify content.
Sanitization: No explicit sanitization or escaping of the interpolated variables is implemented, relying on the agent's internal reasoning and isolation to prevent malicious instructions in the content from compromising the parent agent.

santa-method