The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill heavily interpolates the $ARGUMENTS and $SEVERITY variables directly into natural language prompts for 13 different sub-agent calls across multiple phases. There are no boundary markers (like XML tags or delimiters) or instructions for sub-agents to ignore embedded commands within these variables.
Ingestion points: The $ARGUMENTS variable in SKILL.md is sourced from user input and propagated through the entire workflow.
Boundary markers: Absent. The user-provided string is concatenated directly into urgent instructions (e.g., URGENT: Detect and classify incident: $ARGUMENTS).
Capability inventory: The skill coordinates sub-agents with high-privilege capabilities including deployment execution, system rollbacks, traffic rerouting, and scaling infrastructure.
Sanitization: None present. The raw input is passed to agents responsible for sensitive tasks like Fix Implementation and Security Assessment.
[COMMAND_EXECUTION]: The skill provides instructions to perform high-impact production operations including executing emergency deployments, performing service rollbacks, and modifying traffic routing (throttling/rerouting). While these are intended for incident response, they represent significant control over the production environment.
[DATA_EXFILTRATION]: The workflow involves querying and aggregating highly sensitive telemetry and security data, including distributed traces, log aggregation (ELK/Splunk), APM data, WAF logs, security groups, and audit trails. While no external exfiltration to unauthorized domains is observed, the data processed is highly sensitive.

incident-response-incident-response