The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill explicitly recognizes that telemetry data like $host and $pathname are attacker-controllable and could contain malicious payloads intended for prompt injection.
Ingestion Points: Untrusted data enters the agent context through execute-sql queries on the PostHog events table (specifically $web_vitals events).
Boundary Markers: The skill mandates the use of sanitization logic rather than simple delimiters to create a boundary between raw telemetry and the agent context.
Capability Inventory: The skill is capable of reading data (execute-sql) and writing findings to reports (emit_report, edit_report).
Sanitization: The skill mandates a specific defensive pattern in every SQL query using replaceRegexpAll to strip non-URL-safe characters and substring to cap the length of untrusted strings before they are processed by the LLM or included in reports.
[SAFE]: No other malicious patterns, obfuscation, or unauthorized network operations were detected. The skill is designed to operate within authorized PostHog MCP scopes for analytical purposes.

signals-scout-web-vitals